| REVERIE Instruction: Go to the bathroom with the two towels and empty the lower cabinet. |
| Scene-Oriented Dynamic Planning | w/o Dynamic Demonstration | w/o ROASP |
| Step 1: exit the bedroom | Step 1: put the towel in the bottom drawer of the toilet | Step 1: exit the bathroom |
| Step 2: and walk into the next bedroom in the hall | Step 2: walk into the kitchen |
| Step 3: go stand in the kitchen doorway |
| Step 3: go stand in the bathroom | Step 4: go back to the bedroom and take the towel |
| Step 5: exit the bedroom |
| Step 6: walk into the next bedroom in the hall |
| REVERIE Instruction: Go to the kitchen on level 2 and open the drawer. |
| Scene-Oriented Dynamic Planning | w/o Dynamic Demonstration | w/o ROASP |
| Step 1: go upstairs | Step 1: open the drawer with the key in it | Step 1: go upstairs and turn left at the hall |
| Step 2: go straight pass the couch | Step 2: remove the lock from the door | Step 2: go back down the stairs |
| Step 3: make a slight left towards the refrigerator | Step 3: then turn on | Step 3: go up the stairs again |
| Step 4: wait near the kitchen drawer | Step 4: go back to the kitchen and open the fridge |
| Step 5: take out the fruit and put it in the juicer |
",
+ "bbox": [
+ 95,
+ 130,
+ 875,
+ 349
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Figure 6: Examples of generated instructions.",
+ "bbox": [
+ 332,
+ 359,
+ 633,
+ 372
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "fine-grained static instruction and scene-oriented dynamic instruction is that static fine-grained instruction is generated without ROASP. More specifically, the query prompt for the LLM to generate step-by-step planning is fixed at each timestep, which only consists of the given high-level instruction and the selected demonstration. (III) Dynamic: The input assembled instruction contains the high-level instruction in REVERIE and scene-oriented dynamic planning instruction. As shown in Table 3, in the static setting, the performance in all metrics is improved compared to the baseline, indicating the effectiveness of the LLM's rich world knowledge in fine-grained planning. In the dynamic setting, the performance is further improved with non-trivial margins, showing the effectiveness of ROASP.",
+ "bbox": [
+ 75,
+ 381,
+ 472,
+ 574
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Qualitative Analysis of Prompt Setting To further evaluate the effect of dynamic demonstration and ROASP in SODP, we perform a human evaluation about the generated plannings (see Table 4) and show the planning results (see Fig. 6). For human evaluation, we randomly selected 100 REVERIE tasks and generate fine-grained step-by-step instructions in setting of SODP, SODP without dynamic demonstration, and SODP without ROASP. We asked 10 volunteers to mark the generated step-by-step instructions in terms of their relevancy and rationality. The relevancy score ranges from 0 (unrelated) to 3 (very related), which takes into account whether the keywords in instructions are related to the REVERIE task. For example, regarding the REVERIE instruction \"Go to the kitchen and turn on the microwave\", whether there are keywords in instructions related to the kitchen scene could be rated. Rationality is rated from 0 (bad) to 3 (perfect), considering whether the instruction conforms to the logic of navigation.",
+ "bbox": [
+ 75,
+ 590,
+ 468,
+ 838
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "The results are presented in Table 4. It shows that our SODP scored 2.06 on Relevancy and 1.93 on Rational-",
+ "bbox": [
+ 75,
+ 838,
+ 468,
+ 866
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "ity, which could be considered acceptable since the highest score is 3 and it is challenging to generate instructions that are consistent with tasks and actual navigation logic. When removing the dynamic demonstration, the score of generated instruction drops about $31.55\\%$ on Relevancy and $36.27\\%$ on Rationality, which could also be observed in Fig. 6. Although the instruction generated without dynamic demonstration is related to the task to some extent (e.g., \"put the towel in the bottom drawer of the toilet\" has the keyword \"towel\", the instruction lacks navigation information, such as how to reach the bathroom.) As shown in the bottom example of Fig. 6, instruction without ROASP successfully guided how to go to the destination location kitchen, but it still caused confusion by going upstairs and going downstairs several times, and thus reducing the rationality score, i.e. 1.64 on Relevancy and 1.55 on Rationality. More generation results can be found in the supplementary.",
+ "bbox": [
+ 497,
+ 381,
+ 890,
+ 616
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "6. Conclusions",
+ "text_level": 1,
+ "bbox": [
+ 499,
+ 627,
+ 626,
+ 642
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "In this work, we propose a novel model, March-in-Chat (MiC), for the REVEIRE task, which only provides concise high-level instructions for the VLN agent. MiC enables the REVERIE agent to talk with an LLM on the fly to generate plans for the next few steps. It consists of three main modules, Goal-Oriented Static Planning (GOSP), Scene-Oriented Dynamic Planning (SODP), and Room-and-Object Aware Scene Perceiver (ROASP) module. We conduct extensive quantitative and qualitative experiments on REVERIE and the promising results show the effectiveness of our method.",
+ "bbox": [
+ 497,
+ 650,
+ 890,
+ 802
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "7. Acknowledgements",
+ "text_level": 1,
+ "bbox": [
+ 499,
+ 815,
+ 675,
+ 831
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Jing Liu is supported by the National Key Research and Development Program of China (No. 2020AAA0106400).",
+ "bbox": [
+ 499,
+ 838,
+ 890,
+ 866
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "page_number",
+ "text": "15765",
+ "bbox": [
+ 465,
+ 885,
+ 504,
+ 897
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "References",
+ "text_level": 1,
+ "bbox": [
+ 78,
+ 124,
+ 173,
+ 138
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Daniel Ho, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Eric Jang, Rosario Jauregui Ruano, Kyle Jeffrey, Sally Jesmonth, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Kuang-Huei Lee, Sergey Levine, Yao Lu, Linda Luu, Carolina Parada, Peter Pastor, Jornell Quiambao, Kanishka Rao, Jarek Rettinghouse, Diego Reyes, Pierre Sermanet, Nicolas Sievers, Clayton Tan, Alexander Toshev, Vincent Vanhoucke, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Mengyuan Yan, and Andy Zeng. Do as i can and not as i say: Grounding language in robotic affordances. In arXiv preprint arXiv:2204.01691, 2022. 3",
+ "[2] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Topo-metric map pre-training for language-guided navigation. arXiv preprint arXiv:2212.04385, 2022. 2",
+ "[3] Peter Anderson, Angel X. Chang, Devendra Singh Chaplot, Alexey Dosovitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, and Amir Roshan Zamir. On evaluation of embodied navigation agents. CoRR, abs/1807.06757, 2018. 6",
+ "[4] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, pages 3674-3683, 2018. 1, 2, 5",
+ "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 3, 4",
+ "[6] Angel X. Chang, Angela Dai, Thomas A. Funkhouser, Macej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from RGB-D data in indoor environments. In 3DV, pages 667-676, 2017. 5",
+ "[7] Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. TOUCHDOWN: natural language navigation and spatial reasoning in visual street environments. In CVPR, pages 12538-12547, 2019. 1, 2",
+ "[8] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021. 3",
+ "[9] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, pages 5834-5847, 2021. 7"
+ ],
+ "bbox": [
+ 85,
+ 147,
+ 468,
+ 865
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[10] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Learning from unlabeled 3d environments for vision-and-language navigation. In ECCV, 2022. 4, 6, 7",
+ "[11] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, 2022. 2, 7",
+ "[12] Chen Gao, Jinyu Chen, Si Liu, Luting Wang, Qiong Zhang, and Qi Wu. Room-and-object aware knowledge reasoning for remote embodied referring expression. In CVPR, pages 3064-3073, 2021. 7",
+ "[13] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In ICCV, pages 1634-1643, 2021. 7",
+ "[14] Keji He, Yan Huang, Qi Wu, Jianhua Yang, Dong An, Shuanglin Sima, and Liang Wang. Landmark-rxr: Solving vision-and-language navigation with fine-grained alignment supervision. NeurIPS, 34:652-663, 2021. 2",
+ "[15] Yicong Hong, Cristian Rodriguez Opazo, Qi Wu, and Stephen Gould. Sub-instruction aware vision-and-language navigation. In Bonnie Webber, Trevor Cohn, Yulan He, and Yang Liu, editors, EMNLP, pages 3360-3376, 2020. 5",
+ "[16] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez Opazo, and Stephen Gould. VLN $\\text{©}$ BERT: A recurrent vision-and-language BERT for navigation. In CVPR, pages 1643-1653, 2021. 2, 7",
+ "[17] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In ICML, pages 9118-9147. PMLR, 2022. 2, 3, 5",
+ "[18] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, Pierre Sermanet, Noah Brown, Tomas Jackson, Linda Luu, Sergey Levine, Karol Hausman, and Brian Ichter. Inner monologue: Embodied reasoning through planning with language models. In arXiv preprint arXiv:2207.05608, 2022. 3",
+ "[19] Vihan Jain, Gabriel Magalhães, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. pages 1862-1872, 2019. 2",
+ "[20] Melvin Johnson, Mike Schuster, Quoc V. Le, Maxim Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Thorat, Fernanda B. Viégas, Martin Wattenberg, Greg Corrado, Macduff Hughes, and Jeffrey Dean. Google's multilingual neural machine translation system: Enabling zero-shot translation. Trans. Assoc. Comput. Linguistics, 5:339-351, 2017. 6",
+ "[21] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In EMNLP, pages 4392-4412, 2020. 1, 2",
+ "[22] Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. What makes good in-context examples for gpt-3? arXiv preprint arXiv:2101.06804, 2021. 5"
+ ],
+ "bbox": [
+ 502,
+ 127,
+ 890,
+ 865
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "page_number",
+ "text": "15766",
+ "bbox": [
+ 465,
+ 887,
+ 504,
+ 897
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[23] Chih-Yao Ma, Jiasen Lu, Zuxuan Wu, Ghassan AlRegib, Zsolt Kira, Richard Socher, and Caiming Xiong. Self-monitoring navigation agent via auxiliary progress estimation. In ICLR, 2019. 7",
+ "[24] Khanh Nguyen and Hal Daumé III. Help, anna! visual navigation with natural multimodal assistance via retrospective curiosity-encouraging imitation learning. pages 684–695. Association for Computational Linguistics, 2019. 2",
+ "[25] Gabriel Poesia, Alex Polozov, Vu Le, Ashish Tiwari, Gustavo Soares, Christopher Meek, and Sumit Gulwani. Synchromesh: Reliable code generation from pre-trained language models. In ICLR, 2022. 5",
+ "[26] Yuankai Qi, Zizheng Pan, Yicong Hong, Ming-Hsuan Yang, Anton van den Hengel, and Qi Wu. The road to knowwhere: An object-and-room informed sequential bert for indoor vision-language navigation. In ICCV, pages 1655-1664, 2021. 2, 7",
+ "[27] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. REVERIE: remote embodied visual referring expression in real indoor environments. In CVPR, pages 9979-9988, 2020. 1, 2, 6, 7",
+ "[28] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. HOP: history-and-order aware pretraining for vision-and-language navigation. In CVPR, pages 15397-15406, 2022. 2, 7",
+ "[29] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. Hop+: History-enhanced and order-aware pre-training for vision-and-language navigation. IEEE TPAMI, 2023. 2",
+ "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, volume 139, pages 8748-8763, 2021. 5",
+ "[31] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3, 6",
+ "[32] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors, EMNLP-IJCNLP, pages 3980-3990, 2019. 5",
+ "[33] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A benchmark for interpreting grounded instructions for everyday tasks. In CVPR, pages 10737-10746, 2020. 1, 3",
+ "[34] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. ProgPrompt: Generating situated robot task plans using large language models. 2022. 2",
+ "[35] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. arXiv preprint arXiv:2212.04088, 2022. 3, 5"
+ ],
+ "bbox": [
+ 78,
+ 125,
+ 468,
+ 866
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[36] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In CoRL, pages 394-406, 2019. 1, 2",
+ "[37] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In CVPR, pages 6629-6638, 2019. 7",
+ "[38] Yusheng Zhao, Jinyu Chen, Chen Gao, Wenguan Wang, Lirong Yang, Haibing Ren, Huaxia Xia, and Si Liu. Target-driven structured transformer planner for vision-language navigation. pages 4194-4203, 2022. 7",
+ "[39] Fengda Zhu, Xiwen Liang, Yi Zhu, Qizhi Yu, Xiaojun Chang, and Xiaodan Liang. SOON: scenario oriented object navigation with graph-based exploration. In CVPR, pages 12689-12699, 2021. 1, 2"
+ ],
+ "bbox": [
+ 502,
+ 125,
+ 890,
+ 332
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "page_number",
+ "text": "15767",
+ "bbox": [
+ 465,
+ 885,
+ 504,
+ 897
+ ],
+ "page_idx": 9
+ }
+]
\ No newline at end of file
diff --git a/2023/March in Chat_ Interactive Prompting for Remote Embodied Referring Expression/dfdb2c34-61fe-4960-a099-a29f68122739_model.json b/2023/March in Chat_ Interactive Prompting for Remote Embodied Referring Expression/dfdb2c34-61fe-4960-a099-a29f68122739_model.json
new file mode 100644
index 0000000000000000000000000000000000000000..1fbc8a5b3340e18648c02443524685ad96859b5c
--- /dev/null
+++ b/2023/March in Chat_ Interactive Prompting for Remote Embodied Referring Expression/dfdb2c34-61fe-4960-a099-a29f68122739_model.json
@@ -0,0 +1,2255 @@
+[
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.13,
+ 0.051,
+ 0.201,
+ 0.086
+ ],
+ "angle": 0,
+ "content": "CVF"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.254,
+ 0.049,
+ 0.79,
+ 0.087
+ ],
+ "angle": 0,
+ "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.273,
+ 0.162,
+ 0.7,
+ 0.203
+ ],
+ "angle": 0,
+ "content": "March in Chat: Interactive Prompting for Remote Embodied Referring Expression"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.212,
+ 0.228,
+ 0.764,
+ 0.246
+ ],
+ "angle": 0,
+ "content": "Yanyuan Qiao\\(^{1}\\) Yuankai Qi\\(^{1}\\) Zheng Yu\\(^{1}\\) Jing Liu\\(^{2,3}\\) Qi Wu\\(^{1*}\\)"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.207,
+ 0.246,
+ 0.765,
+ 0.262
+ ],
+ "angle": 0,
+ "content": "\\(^{1}\\)Australian Institute for Machine Learning, The University of Adelaide"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.264,
+ 0.262,
+ 0.706,
+ 0.277
+ ],
+ "angle": 0,
+ "content": "\\(^{2}\\)Institute of Automation, Chinese Academy of Sciences"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.185,
+ 0.277,
+ 0.799,
+ 0.294
+ ],
+ "angle": 0,
+ "content": "\\(^{3}\\)School of Artificial Intelligence, University of Chinese Academy of Sciences"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.296,
+ 0.855,
+ 0.31
+ ],
+ "angle": 0,
+ "content": "{yanyuan.qiao,zheng.yu,qi.wu01}@adelaide.edu.au, qykshr@gmail.com,jliu@nlpr.ia.ac.cn"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.235,
+ 0.342,
+ 0.314,
+ 0.356
+ ],
+ "angle": 0,
+ "content": "Abstract"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.372,
+ 0.474,
+ 0.691
+ ],
+ "angle": 0,
+ "content": "Many Vision-and-Language Navigation (VLN) tasks have been proposed in recent years, from room-based to object-based and indoor to outdoor. The REVERIE (Remote Embodied Referring Expression) is interesting since it only provides high-level instructions to the agent, which are closer to human commands in practice. Nevertheless, this poses more challenges than other VLN tasks since it requires agents to infer a navigation plan only based on a short instruction. Large Language Models (LLMs) show great potential in robot action planning by providing proper prompts. Still, this strategy has not been explored under the REVERIE settings. There are several new challenges. For example, the LLM should be environment-aware so that the navigation plan can be adjusted based on the current visual observation. Moreover, the LLM planned actions should be adaptable to the much larger and more complex REVERIE environment. This paper proposes a March-in-Chat (MiC) model that can talk to the LLM on the fly and plan dynamically based on a newly proposed Room-and-Object Aware Scene Perceiver (ROASP). Our MiC model outperforms the previous state-of-the-art by large margins by SPL and RGSPL metrics on the REVERIE benchmark. The source code is available at https://github.com/YanyuanQiao/MiC"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.078,
+ 0.72,
+ 0.21,
+ 0.734
+ ],
+ "angle": 0,
+ "content": "1. Introduction"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.744,
+ 0.47,
+ 0.842
+ ],
+ "angle": 0,
+ "content": "Vision-and-Language Navigation (VLN), which lies at the intersection of computer vision, natural language processing and robotics, has aroused great attention from research communities in the past few years. Given instructions in natural language, the VLN agent should navigate to the target location based on the dynamic observations in the 3D simulated environments. Since VLN has great"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.513,
+ 0.346,
+ 0.868,
+ 0.369
+ ],
+ "angle": 0,
+ "content": "REVERIE Instruction: Clean the small picture in front of the large mirror."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.52,
+ 0.371,
+ 0.868,
+ 0.515
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.498,
+ 0.526,
+ 0.893,
+ 0.582
+ ],
+ "angle": 0,
+ "content": "Figure 1: Our March-in-Chat (MiC) model is talking to a Large Language Model (LLM) to generate navigation plans on the fly, with the REVERIE instruction and the dynamic room-and-object information as inputs."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.59,
+ 0.892,
+ 0.742
+ ],
+ "angle": 0,
+ "content": "potential in real-world applications such as domestic assistant robots, a large amount of specific VLN tasks have been proposed, including R2R [4] and RxR [21] that ask the agent to navigate from one room to another in a photorealistic environment according to the fine-grained instruction, NDH [36] provides detailed dialogues which imply the instruction, TouchDown [7] extends the task into an outdoor environment, REVERIE [27] and SOON [39] that additionally require the agent's ability of remote object grounding and ALFRED [33] that asks the agent to interact with the target object in a single room of the synthetic environment."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.742,
+ 0.893,
+ 0.867
+ ],
+ "angle": 0,
+ "content": "Most of these VLN tasks provide detailed step-by-step instructions to the agent, such as \"Go up the stairs and then walk the length of the couch. Walk past the dining area and into the kitchen. Stop in front of the refrigerator.\" in R2R. Although detailed instructions can help the agent better achieve the navigation goal in the simulated environments, it has a big gap towards real applications where human beings tend to give coarse-grained high-level instructions such as \"Go to the refrigerator on the second floor\"."
+ },
+ {
+ "type": "page_footnote",
+ "bbox": [
+ 0.096,
+ 0.854,
+ 0.221,
+ 0.867
+ ],
+ "angle": 0,
+ "content": "*Corresponding author"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.466,
+ 0.887,
+ 0.506,
+ 0.898
+ ],
+ "angle": 0,
+ "content": "15758"
+ }
+ ],
+ [
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.126,
+ 0.47,
+ 0.252
+ ],
+ "angle": 0,
+ "content": "Contrary to other tasks, the Remote Embodied Referring Expression (REVERIE) task is more likely to empower the real-world applications of VLN, of which the instructions are closer to those in practice, such as \"Empty the washing machine on level one\". Such high-level instruction is more challenging for VLN agents since it requires them to be more competent in perceiving the surrounding environment and the navigation progress and correspondingly making reasonable plans for the next steps."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.253,
+ 0.471,
+ 0.406
+ ],
+ "angle": 0,
+ "content": "Recently, Large Language Models (LLMs) that internalize a wealth of commonsense knowledge show great potential in action planning for some embodied tasks with the help of suitable in-context learning. However, previous works mainly utilize LLMs to plan atomic actions of object manipulation in a very limited space with simple scenes. These predefined atomic actions can be easily planned well by the LLMs planners with a unified template. Different from these embodied tasks, REVERIE requires large-area exploration from one room to another, which is complex in the layout of rooms and scenes with diverse objects."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.408,
+ 0.471,
+ 0.699
+ ],
+ "angle": 0,
+ "content": "In this work, to adapt LLMs as the planner for REVERIE with the ability of comprehensive scene perception, we propose a novel model named March in Chat (MiC), which enables the LLM as an environment-aware instruction planner through on-the-fly dialogues between the agent and the LLM as Fig. 1 shows. Specifically, the agent is initially situated at the starting position given a high-level coarse-grained REVERIE instruction. First, a Goal-Oriented Static Planning (GOSP) module queries the LLM to point out the target object and infer where the thing may be by using the rich world knowledge internalized in the LLM. Secondly, the agent's Room-and-Object Aware Scene Perceiver (ROASP) describes the current observation and asks the LLM to generate step-by-step fine-grained planning for the next navigation steps. Then, if the ROASP finds the room has changed, the LLM is queried again by the Scene-Oriented Dynamic Planning (SODP) module to generate a new fine-grained step-by-step planning, which will be concatenated with all previous responses from the LLM. The agent will march under the guidance of such interactive prompting until the task is finished."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.701,
+ 0.47,
+ 0.868
+ ],
+ "angle": 0,
+ "content": "To evaluate our proposed MiC, we conduct experiments on the REVERIE benchmark. Our MiC achieves a new state-of-the-art performance in all metrics on REVERIE val unseen set and REVERIE test unseen set. Mainly, MiC obtains \\(41.97\\%\\) on the primary navigation metric of SPL and \\(26.17\\%\\) on the major object grounding metric of RGSPL on test split, which is at least \\(3.09\\%\\) and \\(3.49\\%\\) higher than the previous SoTA results. We also conduct ablation studies to validate the contributions of different components in MiC and the effect of scene-aware perception in dynamic planning generation. These promising results demonstrate the effectiveness of our proposed MiC."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.518,
+ 0.127,
+ 0.854,
+ 0.139
+ ],
+ "angle": 0,
+ "content": "In summary, we make the following contributions:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.518,
+ 0.14,
+ 0.891,
+ 0.181
+ ],
+ "angle": 0,
+ "content": "- We propose a novel March-in-Chat (MiC) model, which lets the REVERIE agent talk with an LLM on the fly to make plans for the next few steps."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.518,
+ 0.183,
+ 0.892,
+ 0.252
+ ],
+ "angle": 0,
+ "content": "- Two planning modules, namely Goal-Oriented Static Planning (GOSP) module, and Scene-Oriented Dynamic Planning (SODP) module, and one Room-and-Object Aware Scene Perceiver (ROASP) module, are proposed."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.518,
+ 0.253,
+ 0.891,
+ 0.293
+ ],
+ "angle": 0,
+ "content": "- Extensive quantitative and qualitative experiments are conducted on REVERIE to validate the effectiveness of our method."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.518,
+ 0.14,
+ 0.892,
+ 0.293
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.305,
+ 0.637,
+ 0.319
+ ],
+ "angle": 0,
+ "content": "2. Related work"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.328,
+ 0.892,
+ 0.659
+ ],
+ "angle": 0,
+ "content": "Vision-and-Language Navigation Vision-and-Language Navigation (VLN) has attracted increasing attention in recent years, and many specific VLN tasks have been proposed [7, 19, 24, 27, 29, 36]. Anderson et al. [4] proposes the first VLN benchmark, Room-to-Room (R2R), which requires an agent to navigate from one room to another in a house, according to a detailed natural language step-by-step instruction in a photo-realistic environment. Later, Room-across-Room (RxR) [21] was proposed with longer and more detailed multilingual instructions. Both these two tasks give fine-grained instructions, which makes it easier to navigate to the target location. NDH [36] extends the navigation instruction to the dialogue form, and TouchDown [7] extends the environments to outdoor. REVERIE [27] and SOON [39] are proposed for remote object localization, which requires an agent not only to navigate to the target location, but also to specify the object to interact with. The difference between REVERIE and SOON is that REVERIE uses short concise instructions (e.g., \"bring me the red cup from the kitchen.\") while SOON employs long detailed instructions (e.g., \"I want to find a cylindrical, metallic and tall lamp, which is set in the bright living room. The lamp is on the cabinet which is on the left of the television and next to the window. The living room is on ...\")."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.66,
+ 0.892,
+ 0.839
+ ],
+ "angle": 0,
+ "content": "Among the aforementioned VLN tasks, the instructions of REVERIE are closer to what we would say to an intelligent domestic robot in daily life in terms of the instruction length and logic, which is usually short and concise. However, most existing methods [2, 11, 14, 16, 26, 28] are usually designed for the VLN tasks where detailed step-by-step instructions are used, thus they do not perform well on REVERIE. In this work, we specifically develop a method for REVERIE. Inspired by the fact of LLMs that implicitly internalize rich knowledge in action planning, we propose to exploit LLM as a fine-grained planner to generate detailed navigation plans from the concise instructions of REVERIE to improve navigation success."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.5,
+ 0.839,
+ 0.892,
+ 0.867
+ ],
+ "angle": 0,
+ "content": "LLMs as Embodied Planner Benefiting from the rise of LLMs, recent works [17, 34] have explored the use of LLMs"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.466,
+ 0.887,
+ 0.506,
+ 0.898
+ ],
+ "angle": 0,
+ "content": "15759"
+ }
+ ],
+ [
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.092,
+ 0.135,
+ 0.466,
+ 0.148
+ ],
+ "angle": 0,
+ "content": "REVERIE instruction: Empty the washing machine on level one."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.524,
+ 0.14,
+ 0.539,
+ 0.149
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.587,
+ 0.14,
+ 0.617,
+ 0.149
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.787,
+ 0.14,
+ 0.803,
+ 0.149
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.092,
+ 0.151,
+ 0.874,
+ 0.319
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.076,
+ 0.329,
+ 0.893,
+ 0.414
+ ],
+ "angle": 0,
+ "content": "Figure 2: Overview of our March-in-Chat model. The program runs along the vertical arrows from left to right, progressing with the time flow. Our model first performs Goal-Oriented Static Planning (GOSP, Sec. 4.1.1) to reason the target object and its possible lying room; then the Room-and-Object Aware Scene Perceiver (ROASP, Sec. 4.2) perceives what room type the agent currently stands in and what prominent objects can be seen; these information are used by the Scene-Oriented Dynamic Planning module (SODP, Sec. 4.1.2) to generate a detailed instruction to execute. GOSP just runs once, and we repeat ROASP and SODP until the agent chooses to stop or reaches the maximum steps."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.426,
+ 0.47,
+ 0.619
+ ],
+ "angle": 0,
+ "content": "in task planning for various embodied tasks. Huang et al. [17] propose to utilize the frozen LLMs (e.g., GPT-2 [31], GPT-3 [5] and Codex [8]) to plan actions for the embodied agent with in-context learning [5]. SayCan [1] translates a high-level instruction into a list of candidate low-level actions with a probability, which is then multiplied by a value function for action prediction. These two LLM planners are static, which only generate action plans at the beginning of a task. By contrast, Huang et al. [18] propose to introduce the feedback of action progress, detected objects and human assistance into the LLM planner to re-plan atomic actions. One concurrent work by Song et al. [35] injects the detected objects to re-generate high-level plans with a fixed program pattern for the ALFRED [33] task."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.62,
+ 0.47,
+ 0.759
+ ],
+ "angle": 0,
+ "content": "However, these above-mentioned methods mainly concentrate on planning atomic actions for object manipulation in a very limited space with simple scenes. By contrast, REVERIE has plenty of much larger and more complicated environments: 90 multi-layer buildings of various styles (e.g., office, home, gym, to name a few). To handle the complex scenarios of REVERIE, we propose a Room-and-Object Aware Scene Perceiver module that helps the LLM planner dynamically interact with the environment in the form of a natural language dialogue."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.078,
+ 0.773,
+ 0.227,
+ 0.789
+ ],
+ "angle": 0,
+ "content": "3. Problem Setup"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.797,
+ 0.469,
+ 0.868
+ ],
+ "angle": 0,
+ "content": "In the REVERIE task, given a concise and high-level instruction referring to a remote object, the agent is expected to navigate to the goal location and identify the target object in previously unseen environments. The environment is defined as an undirected graph \\(\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}\\), where"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.425,
+ 0.892,
+ 0.634
+ ],
+ "angle": 0,
+ "content": "\\(\\mathcal{V} = \\{V_i\\}_{i=1}^K\\) denotes \\(K\\) navigable nodes, and \\(\\mathcal{E}\\) denotes connectivity edges. The agent is first placed in a starting node with the initial state \\(s_0\\) and perceives a panorama \\(\\mathcal{R}_t\\) as the visual observation at each time step \\(t\\). The panorama \\(\\mathcal{R}_t\\) is split into \\(n\\) single view images as \\(\\mathcal{R}_t = \\{r_i\\}_{i=1}^n\\). Each single view image \\(r_i\\) is represented by an image feature vector and an orientation feature vector. In addition, the object features \\(\\mathcal{O}_t = \\{o_i\\}_{i=1}^m\\) of \\(m\\) objects are extracted from the panorama view using the annotated object bounding boxes or object detectors. Then, the agent makes a sequence of actions \\(\\langle a_0, \\ldots, a_T \\rangle\\) to reach the target location, where each action is achieved by choosing a navigable node from the candidate list. The agent navigates in the environment until the target object is grounded or the agent reaches the predefined maximum trajectory length."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.648,
+ 0.59,
+ 0.661
+ ],
+ "angle": 0,
+ "content": "4. Method"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.673,
+ 0.892,
+ 0.868
+ ],
+ "angle": 0,
+ "content": "As illustrated in Fig. 2, when initially situated at the starting position and given a concise high-level instruction such as \"Empty the washing machine on level one\", the agent first queries an LLM with the GOSP module (Sec. 4.1.1) to find out the target object \"washing machine\" in the instruction and reason out the potential location \"laundry room\" by using the world knowledge implied in an LLM. Then the ROASP module (Sec. 4.2) extracts the room type and visible objects from the current visual observation to obtain the environmental feedback. With the description of the scene perception, the LLM is queried again by the SODP (Sec. 4.1.2) to generate the next step instruction, which is used to guide the agent to navigate to the target object in the room. The baseline agent is based on"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.466,
+ 0.887,
+ 0.506,
+ 0.898
+ ],
+ "angle": 0,
+ "content": "15760"
+ }
+ ],
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.225,
+ 0.135,
+ 0.383,
+ 0.147
+ ],
+ "angle": 0,
+ "content": "Target Object Recognition"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.152,
+ 0.148,
+ 0.19,
+ 0.157
+ ],
+ "angle": 0,
+ "content": "Task:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.152,
+ 0.158,
+ 0.428,
+ 0.169
+ ],
+ "angle": 0,
+ "content": "Empty the washing machine on level one."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.151,
+ 0.175,
+ 0.19,
+ 0.184
+ ],
+ "angle": 0,
+ "content": "Goal:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.152,
+ 0.186,
+ 0.416,
+ 0.197
+ ],
+ "angle": 0,
+ "content": "The target object is: washing machine"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.225,
+ 0.212,
+ 0.385,
+ 0.223
+ ],
+ "angle": 0,
+ "content": "Target Object Localization"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.147,
+ 0.223,
+ 0.206,
+ 0.234
+ ],
+ "angle": 0,
+ "content": "Example:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.147,
+ 0.235,
+ 0.462,
+ 0.255
+ ],
+ "angle": 0,
+ "content": "Question: Where does a microwave can usually appear in a house?"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.147,
+ 0.256,
+ 0.262,
+ 0.266
+ ],
+ "angle": 0,
+ "content": "Answer: kitchen."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.147,
+ 0.272,
+ 0.448,
+ 0.294
+ ],
+ "angle": 0,
+ "content": "Question: Where does a washing machine can usually appear in a house?"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.147,
+ 0.295,
+ 0.293,
+ 0.305
+ ],
+ "angle": 0,
+ "content": "Answer: laundry room"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.165,
+ 0.312,
+ 0.444,
+ 0.323
+ ],
+ "angle": 0,
+ "content": "(a) Prompt template of Goal-Oriented Static Planning"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.56,
+ 0.134,
+ 0.741,
+ 0.146
+ ],
+ "angle": 0,
+ "content": "Scene-Oriented Dynamic Plan"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.49,
+ 0.157,
+ 0.811,
+ 0.18
+ ],
+ "angle": 0,
+ "content": "At this step, I am in bedroom, I can see bed, lamp, pillow."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.491,
+ 0.19,
+ 0.548,
+ 0.2
+ ],
+ "angle": 0,
+ "content": "Example:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.491,
+ 0.201,
+ 0.748,
+ 0.212
+ ],
+ "angle": 0,
+ "content": "Step 1: go straight and out the room"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.491,
+ 0.212,
+ 0.692,
+ 0.222
+ ],
+ "angle": 0,
+ "content": "Step 2: pass the white table"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.491,
+ 0.223,
+ 0.811,
+ 0.244
+ ],
+ "angle": 0,
+ "content": "Step 3: move into the laundry room across the hall"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.491,
+ 0.244,
+ 0.663,
+ 0.255
+ ],
+ "angle": 0,
+ "content": "Step 4: stop at the sink"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.49,
+ 0.266,
+ 0.81,
+ 0.288
+ ],
+ "angle": 0,
+ "content": "Task: Empty the washing machine on level one. Step 1: exit the bedroom"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.495,
+ 0.312,
+ 0.798,
+ 0.323
+ ],
+ "angle": 0,
+ "content": "(b) Prompt template of Scene-Oriented Dynamic Planning"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.076,
+ 0.327,
+ 0.892,
+ 0.356
+ ],
+ "angle": 0,
+ "content": "Figure 3: Examples of prompting templates for Goal-Oriented Static Planning (a) and Scene-Oriented Dynamic Planning (b). Outputs are marked with green color. And the red denotes the predicted object from Target Object Recognition."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.37,
+ 0.469,
+ 0.398
+ ],
+ "angle": 0,
+ "content": "HM3D-DUET [10], more details can be found in the supplementary material."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.077,
+ 0.4,
+ 0.458,
+ 0.414
+ ],
+ "angle": 0,
+ "content": "4.1. Planning with World Knowledge from LLMs"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.421,
+ 0.469,
+ 0.504
+ ],
+ "angle": 0,
+ "content": "This section illustrates how we utilize in-context learning (ICL) [5] to acquire world knowledge from the LLMs for planning. We first briefly introduce in-context learning. Then, we elaborate the Goal-Oriented Static Planning (GOSP) and Scene-Oriented Dynamic Planning (SODP). Last, we show the demonstration selection process."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.505,
+ 0.469,
+ 0.656
+ ],
+ "angle": 0,
+ "content": "Preliminary: In-context Learning for Planning. In-context learning (ICL) is a paradigm that lets LLMs directly make predictions based on a natural language context without gradient updates [5]. Specifically, under the setting of in-context learning, an LLM is fed a \"prompt\" that usually contains a task description and several demonstrations, and then the LLM generates the required outputs. Both the prompt template and the choice of demonstration examples have an impact on how well ICL performs. In this work, we use two different ICL settings to generate different navigation plans, i.e. the GOSP and SODP module."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.658,
+ 0.469,
+ 0.796
+ ],
+ "angle": 0,
+ "content": "The GOSP aims to identify the target object and infer the target location by arousing the world knowledge contained in an LLM through appropriate prompts. A fixed demonstration example is used for GOSP. While the SODP aims to generate step-by-step planning instructions after observing the dynamic scenes from the environment, which is more complicated than the former. To better generate plannings, we dynamically select the most suitable demonstration examples for SODP and incorporate the environmental feedback as prompts for interactive planning."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.077,
+ 0.807,
+ 0.403,
+ 0.821
+ ],
+ "angle": 0,
+ "content": "4.1.1 Goal-Oriented Static Planning (GOSP)"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.825,
+ 0.469,
+ 0.867
+ ],
+ "angle": 0,
+ "content": "Given a high-level concise instruction, such as \"Empty the washing machine on level one\", an LLM is first asked to generate a goal-oriented static planning instruction: \"Goal:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.37,
+ 0.892,
+ 0.702
+ ],
+ "angle": 0,
+ "content": "The target object is a washing machine. It is usually in a laundry room\", which emphasizes the target object and points out where the target object may lie. As shown in Fig. 3(a), the planning generation mainly consists of two sub-tasks: target object recognition and target object localization, which can be achieved by providing specifically designed prompts for the LLM. To this end, we design the prompts for the former sub-task in the form: \"Task: Empty the washing machine on level one. Goal: The target object is:\". Then the LLM will generate a corresponding answer: \"washing machine\". By contrast, the latter subtask, reasoning out the target location, is more complex because it requires more suitable prompts to arouse the internalized world knowledge in the LLM. To address this problem, we utilize a fixed demonstration example for the LLM, and design the prompts for the latter sub-task in the form: \"Example: Question: Where does a microwave can usually appear in a house? Answer: kitchen. Question: Where does a washing machine can usually appear in a house? Answer: \". Given such prompts, the LLM will generate the corresponding answer \"laundry room\". With these answers, we can easily combine them into the goal-oriented planning format: \"Goal: The target object is a washing machine. It is usually in a laundry room\"."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.499,
+ 0.711,
+ 0.854,
+ 0.725
+ ],
+ "angle": 0,
+ "content": "4.1.2 Scene-Oriented Dynamic Planning (SODP)"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.729,
+ 0.892,
+ 0.867
+ ],
+ "angle": 0,
+ "content": "As is shown in Fig 3(b), the prompt for SODP consists of three parts. The first part is based on the scene perception of room type, such as \"bedroom\", and visible objects, such as \"bed, lamp, pillow\", obtained by the Room-and-Object Aware Scene Perceiver (ROASP, Sec. 4.2). These information are transformed into a natural language description of the current scene in the format of \"At this step, I am in bedroom, I can see bed, lamp, pillow\". The second part is a demonstration of the fine-grained step-by-step instruction, which is selected according to the strategy"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.466,
+ 0.888,
+ 0.504,
+ 0.898
+ ],
+ "angle": 0,
+ "content": "15761"
+ }
+ ],
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.092,
+ 0.13,
+ 0.118,
+ 0.139
+ ],
+ "angle": 0,
+ "content": "R2R"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.092,
+ 0.14,
+ 0.443,
+ 0.16
+ ],
+ "angle": 0,
+ "content": "Go up the stairs and then walk the length of the couch. Walk past the dining area and into the kitchen. Stop in front of the refrigerator."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.092,
+ 0.163,
+ 0.135,
+ 0.171
+ ],
+ "angle": 0,
+ "content": "FGR2R"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.092,
+ 0.173,
+ 0.197,
+ 0.183
+ ],
+ "angle": 0,
+ "content": "Step 1: go up the stair"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.092,
+ 0.183,
+ 0.302,
+ 0.192
+ ],
+ "angle": 0,
+ "content": "Step 2: and then walk the length of the couch"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.092,
+ 0.193,
+ 0.337,
+ 0.202
+ ],
+ "angle": 0,
+ "content": "Step 3: walk past the dining area and into the kitchen"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.092,
+ 0.202,
+ 0.27,
+ 0.212
+ ],
+ "angle": 0,
+ "content": "Step 4: stop in front of the refrigerator"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.092,
+ 0.173,
+ 0.337,
+ 0.212
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.092,
+ 0.221,
+ 0.453,
+ 0.234
+ ],
+ "angle": 0,
+ "content": "Figure 4: Example of instructions in R2R and FGR2R."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.248,
+ 0.469,
+ 0.331
+ ],
+ "angle": 0,
+ "content": "detailed in the next section. The last part is previous instructions, such as \"Task: Empty the washing machine on level one. Step 1:\". All these three parts are concatenated together and then fed into an LLM to generate the fine-grained planning instruction for the next step accordingly, such as \"Exit the bedroom\"."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.077,
+ 0.34,
+ 0.369,
+ 0.353
+ ],
+ "angle": 0,
+ "content": "4.1.3 Dynamic Demonstration Selection"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.357,
+ 0.469,
+ 0.467
+ ],
+ "angle": 0,
+ "content": "Recent works show that providing various demonstration examples to LLMs benefits the in-context learning for different tasks [17, 22, 25]. In light of these findings, to direct the LLMs in generating better fine-grained plannings, we dynamically select the most suitable demonstration example for each specific task in REVERIE as the prompt to generate the environment-aware instruction, contrary to using a single fixed demonstration for all tasks."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.468,
+ 0.469,
+ 0.648
+ ],
+ "angle": 0,
+ "content": "Specifically, we choose the training set of the Fine-Grained R2R dataset (FGR2R) [15] as the demonstration set \\(\\mathcal{D}\\), of which each sample will be used as a demonstration example \\(D_{step}\\). As shown in Fig. 4, FGR2R decomposes each low-level instruction \\(I_{low}\\) of R2R dataset [4] into step-by-step instructions \\(I_{step}\\). Then, given a high-level instruction \\(I_{high}\\) of REVERIE, a proper \\(I_{low}\\) will be selected as the demonstration example \\(D_{step}\\) by a matching algorithm. In particular, we use \\(I_{high}\\) as query \\(Q\\) and each low-level instruction \\(I_{low}\\) as the key \\(K_i\\), both of which are embedded by the Sentence-BERT [32]. The semantic distance score between the two embeddings is calculated by the cosine similarity:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.175,
+ 0.658,
+ 0.469,
+ 0.689
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathrm {s} (Q, K _ {i}) = \\frac {e (Q) \\cdot e (K _ {i})}{\\| e (Q) \\| \\| e (K _ {i}) \\|}, \\tag {1}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.698,
+ 0.469,
+ 0.768
+ ],
+ "angle": 0,
+ "content": "where \\( e(\\cdot) \\) is the embedding function. If \\( K_{i} \\) has the highest similarity score to the given query \\( Q \\), its corresponding step-by-step instruction \\( I_{step} \\) will be selected as the demonstration example \\( D_{step} \\) for the given high-level instruction \\( I_{high} \\)."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.077,
+ 0.776,
+ 0.43,
+ 0.79
+ ],
+ "angle": 0,
+ "content": "4.2. Room-and-Object Aware Scene Perceiver"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.798,
+ 0.469,
+ 0.866
+ ],
+ "angle": 0,
+ "content": "Though the world knowledge acquired from the static LLMs planner could benefit the embodied task promisingly, the static LLMs planner may generate wrong or irrelevant plannings, which misleads the agent. To address this issue, the LLM planner should be aware of and interact with the"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.127,
+ 0.892,
+ 0.36
+ ],
+ "angle": 0,
+ "content": "dynamic observations. In [35], the names of objects obtained from the ground truth or pre-trained detectors have been added to in-context prompts. However, the agents of these works act in a very limited space with simple scenes and monotonous objects. By contrast, REVERIE involves large-area exploration between different floors and rooms, where the scenes are more complex with more diverse objects. Considering these factors, we propose a room-and-object aware scene perceiver (ROASP) for the LLM planner, which predicts not only the room type but also the visible objects of the current location. Rather than using separate classifiers and detectors to individually predict each position's room types and visible object categories, we use CLIP [30] as the proposed room-and-object aware scene perceiver. Thanks to CLIP's strong ability of zero-shot image classification in the open world, the ROASP can well handle these two tasks."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.361,
+ 0.892,
+ 0.485
+ ],
+ "angle": 0,
+ "content": "Specifically, we first fetch the room type labels from the MatterPort3D [6] semantic annotations and the object type labels are extracted from the REVERIE training dataset. They are used to build the codebook for the room categories \\(\\mathbf{C}_{\\mathrm{room}}\\) and the object categories \\(\\mathbf{C}_{\\mathrm{obj}}\\), respectively. Then, at each timestep \\(t\\), the agent perceives the environment and obtains the panoramic visual observation \\(\\mathcal{R}_t = \\{r_i\\}_{i=1}^n\\). For each single-view observation \\(r_i\\) in the panorama, the image feature \\(f_{\\mathrm{r}}\\) is extracted by the CLIP Image Encoder"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.642,
+ 0.496,
+ 0.891,
+ 0.512
+ ],
+ "angle": 0,
+ "content": "\\[\nf _ {\\mathrm {r}} = E _ {\\mathrm {C L I P}} ^ {v} \\left(r _ {i}\\right), \\tag {2}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.522,
+ 0.892,
+ 0.604
+ ],
+ "angle": 0,
+ "content": "where \\( E_{\\mathrm{CLIP}}^v (\\cdot) \\) represents the CLIP Image Encoder. For each room category \\( c_{room} \\) and each object category \\( c_{obj} \\), we respectively construct a text phrase of room \\( T_{\\mathrm{room}} \\) as \"a photo of a \\( \\{c_{room}\\} \\)\" and a text phrase of object \\( T_{\\mathrm{obj}} \\) as \"a photo of a \\( \\{c_{obj}\\} \\)\". Then the text feature is derived through the pretrained CLIP Text Encoder as:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.621,
+ 0.615,
+ 0.891,
+ 0.631
+ ],
+ "angle": 0,
+ "content": "\\[\nf _ {\\text {r o o m}} = E _ {\\text {C L I P}} ^ {t} \\left(T _ {\\text {r o o m}}\\right), \\tag {3}\n\\]"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.631,
+ 0.634,
+ 0.891,
+ 0.65
+ ],
+ "angle": 0,
+ "content": "\\[\nf _ {\\mathrm {o b j}} = E _ {\\mathrm {C L I P}} ^ {t} \\left(T _ {\\mathrm {o b j}}\\right), \\tag {4}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.659,
+ 0.891,
+ 0.729
+ ],
+ "angle": 0,
+ "content": "where \\( E_{\\mathrm{CLIP}}^{t}(\\cdot) \\) represents the CLIP Text Encoder. At last, the similarity score \\( S_{\\mathrm{room}} \\) between the image feature \\( f_{\\mathrm{r}} \\) and the text feature \\( f_{\\mathrm{room}} \\) as well as the similarity score \\( S_{\\mathrm{obj}} \\) between the image feature \\( f_{\\mathrm{r}} \\) and the text feature \\( f_{\\mathrm{obj}} \\) are respectively computed as:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.596,
+ 0.738,
+ 0.891,
+ 0.755
+ ],
+ "angle": 0,
+ "content": "\\[\nS _ {\\text {r o o m}} = \\operatorname {S o f t m a x} \\left(f _ {\\text {r o o m}} \\cdot f _ {\\mathrm {r}} ^ {T}\\right), \\tag {5}\n\\]"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.607,
+ 0.757,
+ 0.891,
+ 0.774
+ ],
+ "angle": 0,
+ "content": "\\[\nS _ {\\mathrm {o b j}} = \\operatorname {S o f t m a x} \\left(f _ {\\mathrm {o b j}} \\cdot f _ {\\mathrm {r}} ^ {T}\\right). \\tag {6}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.784,
+ 0.892,
+ 0.867
+ ],
+ "angle": 0,
+ "content": "Considering that the current environment normally belongs to only one type of room, though the panoramic images have multiple views, the room that the agent is currently centered in should have the largest influence on each view. Thus, we average the predicted room type scores \\( S_{\\mathrm{room}} \\) from multiple views and choose the room type with"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.466,
+ 0.887,
+ 0.505,
+ 0.898
+ ],
+ "angle": 0,
+ "content": "15762"
+ }
+ ],
+ [
+ {
+ "type": "table",
+ "bbox": [
+ 0.088,
+ 0.131,
+ 0.458,
+ 0.219
+ ],
+ "angle": 0,
+ "content": "| REVERIE Instruction: Go to the bathroom with the two towels and empty the lower cabinet. |
| Scene-Oriented Dynamic Planning | w/o Dynamic Demonstration | w/o ROASP |
| Step 1: exit the bedroom | Step 1: put the towel in the bottom drawer of the toilet | Step 1: exit the bathroom |
| Step 2: and walk into the next bedroom in the hall | Step 2: walk into the kitchen |
| Step 3: go stand in the kitchen doorway |
| Step 3: go stand in the bathroom | Step 4: go back to the bedroom and take the towel |
| Step 5: exit the bedroom |
| Step 6: walk into the next bedroom in the hall |
| REVERIE Instruction: Go to the kitchen on level 2 and open the drawer. |
| Scene-Oriented Dynamic Planning | w/o Dynamic Demonstration | w/o ROASP |
| Step 1: go upstairs | Step 1: open the drawer with the key in it | Step 1: go upstairs and turn left at the hall |
| Step 2: go straight pass the couch | Step 2: remove the lock from the door | Step 2: go back down the stairs |
| Step 3: make a slight left towards the refrigerator | Step 3: then turn on | Step 3: go up the stairs again |
| Step 4: wait near the kitchen drawer | Step 4: go back to the kitchen and open the fridge |
| Step 5: take out the fruit and put it in the juicer |
"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.333,
+ 0.36,
+ 0.635,
+ 0.374
+ ],
+ "angle": 0,
+ "content": "Figure 6: Examples of generated instructions."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.382,
+ 0.473,
+ 0.576
+ ],
+ "angle": 0,
+ "content": "fine-grained static instruction and scene-oriented dynamic instruction is that static fine-grained instruction is generated without ROASP. More specifically, the query prompt for the LLM to generate step-by-step planning is fixed at each timestep, which only consists of the given high-level instruction and the selected demonstration. (III) Dynamic: The input assembled instruction contains the high-level instruction in REVERIE and scene-oriented dynamic planning instruction. As shown in Table 3, in the static setting, the performance in all metrics is improved compared to the baseline, indicating the effectiveness of the LLM's rich world knowledge in fine-grained planning. In the dynamic setting, the performance is further improved with non-trivial margins, showing the effectiveness of ROASP."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.591,
+ 0.47,
+ 0.839
+ ],
+ "angle": 0,
+ "content": "Qualitative Analysis of Prompt Setting To further evaluate the effect of dynamic demonstration and ROASP in SODP, we perform a human evaluation about the generated plannings (see Table 4) and show the planning results (see Fig. 6). For human evaluation, we randomly selected 100 REVERIE tasks and generate fine-grained step-by-step instructions in setting of SODP, SODP without dynamic demonstration, and SODP without ROASP. We asked 10 volunteers to mark the generated step-by-step instructions in terms of their relevancy and rationality. The relevancy score ranges from 0 (unrelated) to 3 (very related), which takes into account whether the keywords in instructions are related to the REVERIE task. For example, regarding the REVERIE instruction \"Go to the kitchen and turn on the microwave\", whether there are keywords in instructions related to the kitchen scene could be rated. Rationality is rated from 0 (bad) to 3 (perfect), considering whether the instruction conforms to the logic of navigation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.839,
+ 0.47,
+ 0.867
+ ],
+ "angle": 0,
+ "content": "The results are presented in Table 4. It shows that our SODP scored 2.06 on Relevancy and 1.93 on Rational-"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.382,
+ 0.892,
+ 0.617
+ ],
+ "angle": 0,
+ "content": "ity, which could be considered acceptable since the highest score is 3 and it is challenging to generate instructions that are consistent with tasks and actual navigation logic. When removing the dynamic demonstration, the score of generated instruction drops about \\(31.55\\%\\) on Relevancy and \\(36.27\\%\\) on Rationality, which could also be observed in Fig. 6. Although the instruction generated without dynamic demonstration is related to the task to some extent (e.g., \"put the towel in the bottom drawer of the toilet\" has the keyword \"towel\", the instruction lacks navigation information, such as how to reach the bathroom.) As shown in the bottom example of Fig. 6, instruction without ROASP successfully guided how to go to the destination location kitchen, but it still caused confusion by going upstairs and going downstairs several times, and thus reducing the rationality score, i.e. 1.64 on Relevancy and 1.55 on Rationality. More generation results can be found in the supplementary."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.628,
+ 0.627,
+ 0.643
+ ],
+ "angle": 0,
+ "content": "6. Conclusions"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.652,
+ 0.892,
+ 0.803
+ ],
+ "angle": 0,
+ "content": "In this work, we propose a novel model, March-in-Chat (MiC), for the REVEIRE task, which only provides concise high-level instructions for the VLN agent. MiC enables the REVERIE agent to talk with an LLM on the fly to generate plans for the next few steps. It consists of three main modules, Goal-Oriented Static Planning (GOSP), Scene-Oriented Dynamic Planning (SODP), and Room-and-Object Aware Scene Perceiver (ROASP) module. We conduct extensive quantitative and qualitative experiments on REVERIE and the promising results show the effectiveness of our method."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.816,
+ 0.676,
+ 0.832
+ ],
+ "angle": 0,
+ "content": "7. Acknowledgements"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.5,
+ 0.839,
+ 0.892,
+ 0.867
+ ],
+ "angle": 0,
+ "content": "Jing Liu is supported by the National Key Research and Development Program of China (No. 2020AAA0106400)."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.466,
+ 0.887,
+ 0.505,
+ 0.898
+ ],
+ "angle": 0,
+ "content": "15765"
+ }
+ ],
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.08,
+ 0.125,
+ 0.174,
+ 0.14
+ ],
+ "angle": 0,
+ "content": "References"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.148,
+ 0.47,
+ 0.325
+ ],
+ "angle": 0,
+ "content": "[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Daniel Ho, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Eric Jang, Rosario Jauregui Ruano, Kyle Jeffrey, Sally Jesmonth, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Kuang-Huei Lee, Sergey Levine, Yao Lu, Linda Luu, Carolina Parada, Peter Pastor, Jornell Quiambao, Kanishka Rao, Jarek Rettinghouse, Diego Reyes, Pierre Sermanet, Nicolas Sievers, Clayton Tan, Alexander Toshev, Vincent Vanhoucke, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Mengyuan Yan, and Andy Zeng. Do as i can and not as i say: Grounding language in robotic affordances. In arXiv preprint arXiv:2204.01691, 2022. 3"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.327,
+ 0.47,
+ 0.377
+ ],
+ "angle": 0,
+ "content": "[2] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Topo-metric map pre-training for language-guided navigation. arXiv preprint arXiv:2212.04385, 2022. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.379,
+ 0.469,
+ 0.442
+ ],
+ "angle": 0,
+ "content": "[3] Peter Anderson, Angel X. Chang, Devendra Singh Chaplot, Alexey Dosovitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, and Amir Roshan Zamir. On evaluation of embodied navigation agents. CoRR, abs/1807.06757, 2018. 6"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.444,
+ 0.469,
+ 0.506
+ ],
+ "angle": 0,
+ "content": "[4] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, pages 3674-3683, 2018. 1, 2, 5"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.508,
+ 0.469,
+ 0.634
+ ],
+ "angle": 0,
+ "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 3, 4"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.636,
+ 0.469,
+ 0.698
+ ],
+ "angle": 0,
+ "content": "[6] Angel X. Chang, Angela Dai, Thomas A. Funkhouser, Macej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from RGB-D data in indoor environments. In 3DV, pages 667-676, 2017. 5"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.7,
+ 0.469,
+ 0.75
+ ],
+ "angle": 0,
+ "content": "[7] Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. TOUCHDOWN: natural language navigation and spatial reasoning in visual street environments. In CVPR, pages 12538-12547, 2019. 1, 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.752,
+ 0.469,
+ 0.814
+ ],
+ "angle": 0,
+ "content": "[8] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021. 3"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.816,
+ 0.469,
+ 0.866
+ ],
+ "angle": 0,
+ "content": "[9] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, pages 5834-5847, 2021. 7"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.087,
+ 0.148,
+ 0.47,
+ 0.866
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.128,
+ 0.892,
+ 0.178
+ ],
+ "angle": 0,
+ "content": "[10] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Learning from unlabeled 3d environments for vision-and-language navigation. In ECCV, 2022. 4, 6, 7"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.18,
+ 0.892,
+ 0.23
+ ],
+ "angle": 0,
+ "content": "[11] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, 2022. 2, 7"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.232,
+ 0.892,
+ 0.282
+ ],
+ "angle": 0,
+ "content": "[12] Chen Gao, Jinyu Chen, Si Liu, Luting Wang, Qiong Zhang, and Qi Wu. Room-and-object aware knowledge reasoning for remote embodied referring expression. In CVPR, pages 3064-3073, 2021. 7"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.284,
+ 0.892,
+ 0.333
+ ],
+ "angle": 0,
+ "content": "[13] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In ICCV, pages 1634-1643, 2021. 7"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.335,
+ 0.892,
+ 0.386
+ ],
+ "angle": 0,
+ "content": "[14] Keji He, Yan Huang, Qi Wu, Jianhua Yang, Dong An, Shuanglin Sima, and Liang Wang. Landmark-rxr: Solving vision-and-language navigation with fine-grained alignment supervision. NeurIPS, 34:652-663, 2021. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.388,
+ 0.892,
+ 0.439
+ ],
+ "angle": 0,
+ "content": "[15] Yicong Hong, Cristian Rodriguez Opazo, Qi Wu, and Stephen Gould. Sub-instruction aware vision-and-language navigation. In Bonnie Webber, Trevor Cohn, Yulan He, and Yang Liu, editors, EMNLP, pages 3360-3376, 2020. 5"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.44,
+ 0.892,
+ 0.49
+ ],
+ "angle": 0,
+ "content": "[16] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez Opazo, and Stephen Gould. VLN\\(\\text{©}\\) BERT: A recurrent vision-and-language BERT for navigation. In CVPR, pages 1643-1653, 2021. 2, 7"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.492,
+ 0.892,
+ 0.543
+ ],
+ "angle": 0,
+ "content": "[17] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In ICML, pages 9118-9147. PMLR, 2022. 2, 3, 5"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.545,
+ 0.892,
+ 0.632
+ ],
+ "angle": 0,
+ "content": "[18] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, Pierre Sermanet, Noah Brown, Tomas Jackson, Linda Luu, Sergey Levine, Karol Hausman, and Brian Ichter. Inner monologue: Embodied reasoning through planning with language models. In arXiv preprint arXiv:2207.05608, 2022. 3"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.634,
+ 0.892,
+ 0.685
+ ],
+ "angle": 0,
+ "content": "[19] Vihan Jain, Gabriel Magalhães, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. pages 1862-1872, 2019. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.687,
+ 0.892,
+ 0.762
+ ],
+ "angle": 0,
+ "content": "[20] Melvin Johnson, Mike Schuster, Quoc V. Le, Maxim Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Thorat, Fernanda B. Viégas, Martin Wattenberg, Greg Corrado, Macduff Hughes, and Jeffrey Dean. Google's multilingual neural machine translation system: Enabling zero-shot translation. Trans. Assoc. Comput. Linguistics, 5:339-351, 2017. 6"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.764,
+ 0.892,
+ 0.814
+ ],
+ "angle": 0,
+ "content": "[21] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In EMNLP, pages 4392-4412, 2020. 1, 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.816,
+ 0.892,
+ 0.866
+ ],
+ "angle": 0,
+ "content": "[22] Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. What makes good in-context examples for gpt-3? arXiv preprint arXiv:2101.06804, 2021. 5"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.503,
+ 0.128,
+ 0.892,
+ 0.866
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.467,
+ 0.888,
+ 0.505,
+ 0.898
+ ],
+ "angle": 0,
+ "content": "15766"
+ }
+ ],
+ [
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.127,
+ 0.47,
+ 0.178
+ ],
+ "angle": 0,
+ "content": "[23] Chih-Yao Ma, Jiasen Lu, Zuxuan Wu, Ghassan AlRegib, Zsolt Kira, Richard Socher, and Caiming Xiong. Self-monitoring navigation agent via auxiliary progress estimation. In ICLR, 2019. 7"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.18,
+ 0.469,
+ 0.231
+ ],
+ "angle": 0,
+ "content": "[24] Khanh Nguyen and Hal Daumé III. Help, anna! visual navigation with natural multimodal assistance via retrospective curiosity-encouraging imitation learning. pages 684–695. Association for Computational Linguistics, 2019. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.232,
+ 0.469,
+ 0.282
+ ],
+ "angle": 0,
+ "content": "[25] Gabriel Poesia, Alex Polozov, Vu Le, Ashish Tiwari, Gustavo Soares, Christopher Meek, and Sumit Gulwani. Synchromesh: Reliable code generation from pre-trained language models. In ICLR, 2022. 5"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.284,
+ 0.469,
+ 0.347
+ ],
+ "angle": 0,
+ "content": "[26] Yuankai Qi, Zizheng Pan, Yicong Hong, Ming-Hsuan Yang, Anton van den Hengel, and Qi Wu. The road to knowwhere: An object-and-room informed sequential bert for indoor vision-language navigation. In ICCV, pages 1655-1664, 2021. 2, 7"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.349,
+ 0.469,
+ 0.411
+ ],
+ "angle": 0,
+ "content": "[27] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. REVERIE: remote embodied visual referring expression in real indoor environments. In CVPR, pages 9979-9988, 2020. 1, 2, 6, 7"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.413,
+ 0.469,
+ 0.463
+ ],
+ "angle": 0,
+ "content": "[28] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. HOP: history-and-order aware pretraining for vision-and-language navigation. In CVPR, pages 15397-15406, 2022. 2, 7"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.465,
+ 0.469,
+ 0.515
+ ],
+ "angle": 0,
+ "content": "[29] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. Hop+: History-enhanced and order-aware pre-training for vision-and-language navigation. IEEE TPAMI, 2023. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.517,
+ 0.469,
+ 0.594
+ ],
+ "angle": 0,
+ "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, volume 139, pages 8748-8763, 2021. 5"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.595,
+ 0.469,
+ 0.644
+ ],
+ "angle": 0,
+ "content": "[31] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3, 6"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.647,
+ 0.469,
+ 0.697
+ ],
+ "angle": 0,
+ "content": "[32] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors, EMNLP-IJCNLP, pages 3980-3990, 2019. 5"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.699,
+ 0.469,
+ 0.761
+ ],
+ "angle": 0,
+ "content": "[33] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A benchmark for interpreting grounded instructions for everyday tasks. In CVPR, pages 10737-10746, 2020. 1, 3"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.763,
+ 0.469,
+ 0.814
+ ],
+ "angle": 0,
+ "content": "[34] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. ProgPrompt: Generating situated robot task plans using large language models. 2022. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.816,
+ 0.469,
+ 0.867
+ ],
+ "angle": 0,
+ "content": "[35] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. arXiv preprint arXiv:2212.04088, 2022. 3, 5"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.08,
+ 0.127,
+ 0.47,
+ 0.867
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.127,
+ 0.892,
+ 0.165
+ ],
+ "angle": 0,
+ "content": "[36] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In CoRL, pages 394-406, 2019. 1, 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.167,
+ 0.892,
+ 0.23
+ ],
+ "angle": 0,
+ "content": "[37] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In CVPR, pages 6629-6638, 2019. 7"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.231,
+ 0.892,
+ 0.281
+ ],
+ "angle": 0,
+ "content": "[38] Yusheng Zhao, Jinyu Chen, Chen Gao, Wenguan Wang, Lirong Yang, Haibing Ren, Huaxia Xia, and Si Liu. Target-driven structured transformer planner for vision-language navigation. pages 4194-4203, 2022. 7"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.282,
+ 0.892,
+ 0.333
+ ],
+ "angle": 0,
+ "content": "[39] Fengda Zhu, Xiwen Liang, Yi Zhu, Qizhi Yu, Xiaojun Chang, and Xiaodan Liang. SOON: scenario oriented object navigation with graph-based exploration. In CVPR, pages 12689-12699, 2021. 1, 2"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.503,
+ 0.127,
+ 0.892,
+ 0.333
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.467,
+ 0.887,
+ 0.505,
+ 0.898
+ ],
+ "angle": 0,
+ "content": "15767"
+ }
+ ]
+]
\ No newline at end of file
diff --git a/2023/March in Chat_ Interactive Prompting for Remote Embodied Referring Expression/dfdb2c34-61fe-4960-a099-a29f68122739_origin.pdf b/2023/March in Chat_ Interactive Prompting for Remote Embodied Referring Expression/dfdb2c34-61fe-4960-a099-a29f68122739_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..1eef2513fa45b2efff9ec5f5929598f83ea6f826
--- /dev/null
+++ b/2023/March in Chat_ Interactive Prompting for Remote Embodied Referring Expression/dfdb2c34-61fe-4960-a099-a29f68122739_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c7d4cd53755eaa217772b5f750f86304fb7ec6abe225347a084979b50d2eb2e
+size 492616
diff --git a/2023/March in Chat_ Interactive Prompting for Remote Embodied Referring Expression/full.md b/2023/March in Chat_ Interactive Prompting for Remote Embodied Referring Expression/full.md
new file mode 100644
index 0000000000000000000000000000000000000000..117c1bd59db6c19bfb2c87ed948da712db47dfe5
--- /dev/null
+++ b/2023/March in Chat_ Interactive Prompting for Remote Embodied Referring Expression/full.md
@@ -0,0 +1,330 @@
+# March in Chat: Interactive Prompting for Remote Embodied Referring Expression
+
+Yanyuan Qiao $^{1}$ Yuankai Qi $^{1}$ Zheng Yu $^{1}$ Jing Liu $^{2,3}$ Qi Wu $^{1*}$
+
+$^{1}$ Australian Institute for Machine Learning, The University of Adelaide
+
+$^{2}$ Institute of Automation, Chinese Academy of Sciences
+
+$^{3}$ School of Artificial Intelligence, University of Chinese Academy of Sciences
+
+{yanyuan.qiao,zheng.yu,qi.wu01}@adelaide.edu.au, qykshr@gmail.com,jliu@nlpr.ia.ac.cn
+
+# Abstract
+
+Many Vision-and-Language Navigation (VLN) tasks have been proposed in recent years, from room-based to object-based and indoor to outdoor. The REVERIE (Remote Embodied Referring Expression) is interesting since it only provides high-level instructions to the agent, which are closer to human commands in practice. Nevertheless, this poses more challenges than other VLN tasks since it requires agents to infer a navigation plan only based on a short instruction. Large Language Models (LLMs) show great potential in robot action planning by providing proper prompts. Still, this strategy has not been explored under the REVERIE settings. There are several new challenges. For example, the LLM should be environment-aware so that the navigation plan can be adjusted based on the current visual observation. Moreover, the LLM planned actions should be adaptable to the much larger and more complex REVERIE environment. This paper proposes a March-in-Chat (MiC) model that can talk to the LLM on the fly and plan dynamically based on a newly proposed Room-and-Object Aware Scene Perceiver (ROASP). Our MiC model outperforms the previous state-of-the-art by large margins by SPL and RGSPL metrics on the REVERIE benchmark. The source code is available at https://github.com/YanyuanQiao/MiC
+
+# 1. Introduction
+
+Vision-and-Language Navigation (VLN), which lies at the intersection of computer vision, natural language processing and robotics, has aroused great attention from research communities in the past few years. Given instructions in natural language, the VLN agent should navigate to the target location based on the dynamic observations in the 3D simulated environments. Since VLN has great
+
+
+REVERIE Instruction: Clean the small picture in front of the large mirror.
+Figure 1: Our March-in-Chat (MiC) model is talking to a Large Language Model (LLM) to generate navigation plans on the fly, with the REVERIE instruction and the dynamic room-and-object information as inputs.
+
+potential in real-world applications such as domestic assistant robots, a large amount of specific VLN tasks have been proposed, including R2R [4] and RxR [21] that ask the agent to navigate from one room to another in a photorealistic environment according to the fine-grained instruction, NDH [36] provides detailed dialogues which imply the instruction, TouchDown [7] extends the task into an outdoor environment, REVERIE [27] and SOON [39] that additionally require the agent's ability of remote object grounding and ALFRED [33] that asks the agent to interact with the target object in a single room of the synthetic environment.
+
+Most of these VLN tasks provide detailed step-by-step instructions to the agent, such as "Go up the stairs and then walk the length of the couch. Walk past the dining area and into the kitchen. Stop in front of the refrigerator." in R2R. Although detailed instructions can help the agent better achieve the navigation goal in the simulated environments, it has a big gap towards real applications where human beings tend to give coarse-grained high-level instructions such as "Go to the refrigerator on the second floor".
+
+Contrary to other tasks, the Remote Embodied Referring Expression (REVERIE) task is more likely to empower the real-world applications of VLN, of which the instructions are closer to those in practice, such as "Empty the washing machine on level one". Such high-level instruction is more challenging for VLN agents since it requires them to be more competent in perceiving the surrounding environment and the navigation progress and correspondingly making reasonable plans for the next steps.
+
+Recently, Large Language Models (LLMs) that internalize a wealth of commonsense knowledge show great potential in action planning for some embodied tasks with the help of suitable in-context learning. However, previous works mainly utilize LLMs to plan atomic actions of object manipulation in a very limited space with simple scenes. These predefined atomic actions can be easily planned well by the LLMs planners with a unified template. Different from these embodied tasks, REVERIE requires large-area exploration from one room to another, which is complex in the layout of rooms and scenes with diverse objects.
+
+In this work, to adapt LLMs as the planner for REVERIE with the ability of comprehensive scene perception, we propose a novel model named March in Chat (MiC), which enables the LLM as an environment-aware instruction planner through on-the-fly dialogues between the agent and the LLM as Fig. 1 shows. Specifically, the agent is initially situated at the starting position given a high-level coarse-grained REVERIE instruction. First, a Goal-Oriented Static Planning (GOSP) module queries the LLM to point out the target object and infer where the thing may be by using the rich world knowledge internalized in the LLM. Secondly, the agent's Room-and-Object Aware Scene Perceiver (ROASP) describes the current observation and asks the LLM to generate step-by-step fine-grained planning for the next navigation steps. Then, if the ROASP finds the room has changed, the LLM is queried again by the Scene-Oriented Dynamic Planning (SODP) module to generate a new fine-grained step-by-step planning, which will be concatenated with all previous responses from the LLM. The agent will march under the guidance of such interactive prompting until the task is finished.
+
+To evaluate our proposed MiC, we conduct experiments on the REVERIE benchmark. Our MiC achieves a new state-of-the-art performance in all metrics on REVERIE val unseen set and REVERIE test unseen set. Mainly, MiC obtains $41.97\%$ on the primary navigation metric of SPL and $26.17\%$ on the major object grounding metric of RGSPL on test split, which is at least $3.09\%$ and $3.49\%$ higher than the previous SoTA results. We also conduct ablation studies to validate the contributions of different components in MiC and the effect of scene-aware perception in dynamic planning generation. These promising results demonstrate the effectiveness of our proposed MiC.
+
+In summary, we make the following contributions:
+
+- We propose a novel March-in-Chat (MiC) model, which lets the REVERIE agent talk with an LLM on the fly to make plans for the next few steps.
+- Two planning modules, namely Goal-Oriented Static Planning (GOSP) module, and Scene-Oriented Dynamic Planning (SODP) module, and one Room-and-Object Aware Scene Perceiver (ROASP) module, are proposed.
+- Extensive quantitative and qualitative experiments are conducted on REVERIE to validate the effectiveness of our method.
+
+# 2. Related work
+
+Vision-and-Language Navigation Vision-and-Language Navigation (VLN) has attracted increasing attention in recent years, and many specific VLN tasks have been proposed [7, 19, 24, 27, 29, 36]. Anderson et al. [4] proposes the first VLN benchmark, Room-to-Room (R2R), which requires an agent to navigate from one room to another in a house, according to a detailed natural language step-by-step instruction in a photo-realistic environment. Later, Room-across-Room (RxR) [21] was proposed with longer and more detailed multilingual instructions. Both these two tasks give fine-grained instructions, which makes it easier to navigate to the target location. NDH [36] extends the navigation instruction to the dialogue form, and TouchDown [7] extends the environments to outdoor. REVERIE [27] and SOON [39] are proposed for remote object localization, which requires an agent not only to navigate to the target location, but also to specify the object to interact with. The difference between REVERIE and SOON is that REVERIE uses short concise instructions (e.g., "bring me the red cup from the kitchen.") while SOON employs long detailed instructions (e.g., "I want to find a cylindrical, metallic and tall lamp, which is set in the bright living room. The lamp is on the cabinet which is on the left of the television and next to the window. The living room is on ...").
+
+Among the aforementioned VLN tasks, the instructions of REVERIE are closer to what we would say to an intelligent domestic robot in daily life in terms of the instruction length and logic, which is usually short and concise. However, most existing methods [2, 11, 14, 16, 26, 28] are usually designed for the VLN tasks where detailed step-by-step instructions are used, thus they do not perform well on REVERIE. In this work, we specifically develop a method for REVERIE. Inspired by the fact of LLMs that implicitly internalize rich knowledge in action planning, we propose to exploit LLM as a fine-grained planner to generate detailed navigation plans from the concise instructions of REVERIE to improve navigation success.
+
+LLMs as Embodied Planner Benefiting from the rise of LLMs, recent works [17, 34] have explored the use of LLMs
+
+
+
+
+
+
+
+
+REVERIE instruction: Empty the washing machine on level one.
+Figure 2: Overview of our March-in-Chat model. The program runs along the vertical arrows from left to right, progressing with the time flow. Our model first performs Goal-Oriented Static Planning (GOSP, Sec. 4.1.1) to reason the target object and its possible lying room; then the Room-and-Object Aware Scene Perceiver (ROASP, Sec. 4.2) perceives what room type the agent currently stands in and what prominent objects can be seen; these information are used by the Scene-Oriented Dynamic Planning module (SODP, Sec. 4.1.2) to generate a detailed instruction to execute. GOSP just runs once, and we repeat ROASP and SODP until the agent chooses to stop or reaches the maximum steps.
+
+in task planning for various embodied tasks. Huang et al. [17] propose to utilize the frozen LLMs (e.g., GPT-2 [31], GPT-3 [5] and Codex [8]) to plan actions for the embodied agent with in-context learning [5]. SayCan [1] translates a high-level instruction into a list of candidate low-level actions with a probability, which is then multiplied by a value function for action prediction. These two LLM planners are static, which only generate action plans at the beginning of a task. By contrast, Huang et al. [18] propose to introduce the feedback of action progress, detected objects and human assistance into the LLM planner to re-plan atomic actions. One concurrent work by Song et al. [35] injects the detected objects to re-generate high-level plans with a fixed program pattern for the ALFRED [33] task.
+
+However, these above-mentioned methods mainly concentrate on planning atomic actions for object manipulation in a very limited space with simple scenes. By contrast, REVERIE has plenty of much larger and more complicated environments: 90 multi-layer buildings of various styles (e.g., office, home, gym, to name a few). To handle the complex scenarios of REVERIE, we propose a Room-and-Object Aware Scene Perceiver module that helps the LLM planner dynamically interact with the environment in the form of a natural language dialogue.
+
+# 3. Problem Setup
+
+In the REVERIE task, given a concise and high-level instruction referring to a remote object, the agent is expected to navigate to the goal location and identify the target object in previously unseen environments. The environment is defined as an undirected graph $\mathcal{G} = \{\mathcal{V},\mathcal{E}\}$ , where
+
+$\mathcal{V} = \{V_i\}_{i=1}^K$ denotes $K$ navigable nodes, and $\mathcal{E}$ denotes connectivity edges. The agent is first placed in a starting node with the initial state $s_0$ and perceives a panorama $\mathcal{R}_t$ as the visual observation at each time step $t$ . The panorama $\mathcal{R}_t$ is split into $n$ single view images as $\mathcal{R}_t = \{r_i\}_{i=1}^n$ . Each single view image $r_i$ is represented by an image feature vector and an orientation feature vector. In addition, the object features $\mathcal{O}_t = \{o_i\}_{i=1}^m$ of $m$ objects are extracted from the panorama view using the annotated object bounding boxes or object detectors. Then, the agent makes a sequence of actions $\langle a_0, \ldots, a_T \rangle$ to reach the target location, where each action is achieved by choosing a navigable node from the candidate list. The agent navigates in the environment until the target object is grounded or the agent reaches the predefined maximum trajectory length.
+
+# 4. Method
+
+As illustrated in Fig. 2, when initially situated at the starting position and given a concise high-level instruction such as "Empty the washing machine on level one", the agent first queries an LLM with the GOSP module (Sec. 4.1.1) to find out the target object "washing machine" in the instruction and reason out the potential location "laundry room" by using the world knowledge implied in an LLM. Then the ROASP module (Sec. 4.2) extracts the room type and visible objects from the current visual observation to obtain the environmental feedback. With the description of the scene perception, the LLM is queried again by the SODP (Sec. 4.1.2) to generate the next step instruction, which is used to guide the agent to navigate to the target object in the room. The baseline agent is based on
+
+# Target Object Recognition
+
+Task:
+
+Empty the washing machine on level one.
+
+Goal:
+
+The target object is: washing machine
+
+# Target Object Localization
+
+Example:
+
+Question: Where does a microwave can usually appear in a house?
+
+Answer: kitchen.
+
+Question: Where does a washing machine can usually appear in a house?
+
+Answer: laundry room
+
+(a) Prompt template of Goal-Oriented Static Planning
+
+# Scene-Oriented Dynamic Plan
+
+At this step, I am in bedroom, I can see bed, lamp, pillow.
+
+Example:
+
+Step 1: go straight and out the room
+
+Step 2: pass the white table
+
+Step 3: move into the laundry room across the hall
+
+Step 4: stop at the sink
+
+Task: Empty the washing machine on level one. Step 1: exit the bedroom
+
+(b) Prompt template of Scene-Oriented Dynamic Planning
+
+Figure 3: Examples of prompting templates for Goal-Oriented Static Planning (a) and Scene-Oriented Dynamic Planning (b). Outputs are marked with green color. And the red denotes the predicted object from Target Object Recognition.
+
+HM3D-DUET [10], more details can be found in the supplementary material.
+
+# 4.1. Planning with World Knowledge from LLMs
+
+This section illustrates how we utilize in-context learning (ICL) [5] to acquire world knowledge from the LLMs for planning. We first briefly introduce in-context learning. Then, we elaborate the Goal-Oriented Static Planning (GOSP) and Scene-Oriented Dynamic Planning (SODP). Last, we show the demonstration selection process.
+
+Preliminary: In-context Learning for Planning. In-context learning (ICL) is a paradigm that lets LLMs directly make predictions based on a natural language context without gradient updates [5]. Specifically, under the setting of in-context learning, an LLM is fed a "prompt" that usually contains a task description and several demonstrations, and then the LLM generates the required outputs. Both the prompt template and the choice of demonstration examples have an impact on how well ICL performs. In this work, we use two different ICL settings to generate different navigation plans, i.e. the GOSP and SODP module.
+
+The GOSP aims to identify the target object and infer the target location by arousing the world knowledge contained in an LLM through appropriate prompts. A fixed demonstration example is used for GOSP. While the SODP aims to generate step-by-step planning instructions after observing the dynamic scenes from the environment, which is more complicated than the former. To better generate plannings, we dynamically select the most suitable demonstration examples for SODP and incorporate the environmental feedback as prompts for interactive planning.
+
+# 4.1.1 Goal-Oriented Static Planning (GOSP)
+
+Given a high-level concise instruction, such as "Empty the washing machine on level one", an LLM is first asked to generate a goal-oriented static planning instruction: "Goal:
+
+The target object is a washing machine. It is usually in a laundry room", which emphasizes the target object and points out where the target object may lie. As shown in Fig. 3(a), the planning generation mainly consists of two sub-tasks: target object recognition and target object localization, which can be achieved by providing specifically designed prompts for the LLM. To this end, we design the prompts for the former sub-task in the form: "Task: Empty the washing machine on level one. Goal: The target object is:". Then the LLM will generate a corresponding answer: "washing machine". By contrast, the latter subtask, reasoning out the target location, is more complex because it requires more suitable prompts to arouse the internalized world knowledge in the LLM. To address this problem, we utilize a fixed demonstration example for the LLM, and design the prompts for the latter sub-task in the form: "Example: Question: Where does a microwave can usually appear in a house? Answer: kitchen. Question: Where does a washing machine can usually appear in a house? Answer: ". Given such prompts, the LLM will generate the corresponding answer "laundry room". With these answers, we can easily combine them into the goal-oriented planning format: "Goal: The target object is a washing machine. It is usually in a laundry room".
+
+# 4.1.2 Scene-Oriented Dynamic Planning (SODP)
+
+As is shown in Fig 3(b), the prompt for SODP consists of three parts. The first part is based on the scene perception of room type, such as "bedroom", and visible objects, such as "bed, lamp, pillow", obtained by the Room-and-Object Aware Scene Perceiver (ROASP, Sec. 4.2). These information are transformed into a natural language description of the current scene in the format of "At this step, I am in bedroom, I can see bed, lamp, pillow". The second part is a demonstration of the fine-grained step-by-step instruction, which is selected according to the strategy
+
+# R2R
+
+Go up the stairs and then walk the length of the couch. Walk past the dining area and into the kitchen. Stop in front of the refrigerator.
+
+# FGR2R
+
+Step 1: go up the stair
+Step 2: and then walk the length of the couch
+Step 3: walk past the dining area and into the kitchen
+Step 4: stop in front of the refrigerator
+
+Figure 4: Example of instructions in R2R and FGR2R.
+
+detailed in the next section. The last part is previous instructions, such as "Task: Empty the washing machine on level one. Step 1:". All these three parts are concatenated together and then fed into an LLM to generate the fine-grained planning instruction for the next step accordingly, such as "Exit the bedroom".
+
+# 4.1.3 Dynamic Demonstration Selection
+
+Recent works show that providing various demonstration examples to LLMs benefits the in-context learning for different tasks [17, 22, 25]. In light of these findings, to direct the LLMs in generating better fine-grained plannings, we dynamically select the most suitable demonstration example for each specific task in REVERIE as the prompt to generate the environment-aware instruction, contrary to using a single fixed demonstration for all tasks.
+
+Specifically, we choose the training set of the Fine-Grained R2R dataset (FGR2R) [15] as the demonstration set $\mathcal{D}$ , of which each sample will be used as a demonstration example $D_{step}$ . As shown in Fig. 4, FGR2R decomposes each low-level instruction $I_{low}$ of R2R dataset [4] into step-by-step instructions $I_{step}$ . Then, given a high-level instruction $I_{high}$ of REVERIE, a proper $I_{low}$ will be selected as the demonstration example $D_{step}$ by a matching algorithm. In particular, we use $I_{high}$ as query $Q$ and each low-level instruction $I_{low}$ as the key $K_i$ , both of which are embedded by the Sentence-BERT [32]. The semantic distance score between the two embeddings is calculated by the cosine similarity:
+
+$$
+\mathrm {s} (Q, K _ {i}) = \frac {e (Q) \cdot e (K _ {i})}{\| e (Q) \| \| e (K _ {i}) \|}, \tag {1}
+$$
+
+where $e(\cdot)$ is the embedding function. If $K_{i}$ has the highest similarity score to the given query $Q$ , its corresponding step-by-step instruction $I_{step}$ will be selected as the demonstration example $D_{step}$ for the given high-level instruction $I_{high}$ .
+
+# 4.2. Room-and-Object Aware Scene Perceiver
+
+Though the world knowledge acquired from the static LLMs planner could benefit the embodied task promisingly, the static LLMs planner may generate wrong or irrelevant plannings, which misleads the agent. To address this issue, the LLM planner should be aware of and interact with the
+
+dynamic observations. In [35], the names of objects obtained from the ground truth or pre-trained detectors have been added to in-context prompts. However, the agents of these works act in a very limited space with simple scenes and monotonous objects. By contrast, REVERIE involves large-area exploration between different floors and rooms, where the scenes are more complex with more diverse objects. Considering these factors, we propose a room-and-object aware scene perceiver (ROASP) for the LLM planner, which predicts not only the room type but also the visible objects of the current location. Rather than using separate classifiers and detectors to individually predict each position's room types and visible object categories, we use CLIP [30] as the proposed room-and-object aware scene perceiver. Thanks to CLIP's strong ability of zero-shot image classification in the open world, the ROASP can well handle these two tasks.
+
+Specifically, we first fetch the room type labels from the MatterPort3D [6] semantic annotations and the object type labels are extracted from the REVERIE training dataset. They are used to build the codebook for the room categories $\mathbf{C}_{\mathrm{room}}$ and the object categories $\mathbf{C}_{\mathrm{obj}}$ , respectively. Then, at each timestep $t$ , the agent perceives the environment and obtains the panoramic visual observation $\mathcal{R}_t = \{r_i\}_{i=1}^n$ . For each single-view observation $r_i$ in the panorama, the image feature $f_{\mathrm{r}}$ is extracted by the CLIP Image Encoder
+
+$$
+f _ {\mathrm {r}} = E _ {\mathrm {C L I P}} ^ {v} \left(r _ {i}\right), \tag {2}
+$$
+
+where $E_{\mathrm{CLIP}}^v (\cdot)$ represents the CLIP Image Encoder. For each room category $c_{room}$ and each object category $c_{obj}$ , we respectively construct a text phrase of room $T_{\mathrm{room}}$ as "a photo of a $\{c_{room}\}$ " and a text phrase of object $T_{\mathrm{obj}}$ as "a photo of a $\{c_{obj}\}$ ". Then the text feature is derived through the pretrained CLIP Text Encoder as:
+
+$$
+f _ {\text {r o o m}} = E _ {\text {C L I P}} ^ {t} \left(T _ {\text {r o o m}}\right), \tag {3}
+$$
+
+$$
+f _ {\mathrm {o b j}} = E _ {\mathrm {C L I P}} ^ {t} \left(T _ {\mathrm {o b j}}\right), \tag {4}
+$$
+
+where $E_{\mathrm{CLIP}}^{t}(\cdot)$ represents the CLIP Text Encoder. At last, the similarity score $S_{\mathrm{room}}$ between the image feature $f_{\mathrm{r}}$ and the text feature $f_{\mathrm{room}}$ as well as the similarity score $S_{\mathrm{obj}}$ between the image feature $f_{\mathrm{r}}$ and the text feature $f_{\mathrm{obj}}$ are respectively computed as:
+
+$$
+S _ {\text {r o o m}} = \operatorname {S o f t m a x} \left(f _ {\text {r o o m}} \cdot f _ {\mathrm {r}} ^ {T}\right), \tag {5}
+$$
+
+$$
+S _ {\mathrm {o b j}} = \operatorname {S o f t m a x} \left(f _ {\mathrm {o b j}} \cdot f _ {\mathrm {r}} ^ {T}\right). \tag {6}
+$$
+
+Considering that the current environment normally belongs to only one type of room, though the panoramic images have multiple views, the room that the agent is currently centered in should have the largest influence on each view. Thus, we average the predicted room type scores $S_{\mathrm{room}}$ from multiple views and choose the room type with
+
+| REVERIE Instruction: Go to the bathroom with the two towels and empty the lower cabinet. |
| Scene-Oriented Dynamic Planning | w/o Dynamic Demonstration | w/o ROASP |
| Step 1: exit the bedroom | Step 1: put the towel in the bottom drawer of the toilet | Step 1: exit the bathroom |
| Step 2: and walk into the next bedroom in the hall | Step 2: walk into the kitchen |
| Step 3: go stand in the kitchen doorway |
| Step 3: go stand in the bathroom | Step 4: go back to the bedroom and take the towel |
| Step 5: exit the bedroom |
| Step 6: walk into the next bedroom in the hall |
| REVERIE Instruction: Go to the kitchen on level 2 and open the drawer. |
| Scene-Oriented Dynamic Planning | w/o Dynamic Demonstration | w/o ROASP |
| Step 1: go upstairs | Step 1: open the drawer with the key in it | Step 1: go upstairs and turn left at the hall |
| Step 2: go straight pass the couch | Step 2: remove the lock from the door | Step 2: go back down the stairs |
| Step 3: make a slight left towards the refrigerator | Step 3: then turn on | Step 3: go up the stairs again |
| Step 4: wait near the kitchen drawer | Step 4: go back to the kitchen and open the fridge |
| Step 5: take out the fruit and put it in the juicer |
+
+Figure 6: Examples of generated instructions.
+
+fine-grained static instruction and scene-oriented dynamic instruction is that static fine-grained instruction is generated without ROASP. More specifically, the query prompt for the LLM to generate step-by-step planning is fixed at each timestep, which only consists of the given high-level instruction and the selected demonstration. (III) Dynamic: The input assembled instruction contains the high-level instruction in REVERIE and scene-oriented dynamic planning instruction. As shown in Table 3, in the static setting, the performance in all metrics is improved compared to the baseline, indicating the effectiveness of the LLM's rich world knowledge in fine-grained planning. In the dynamic setting, the performance is further improved with non-trivial margins, showing the effectiveness of ROASP.
+
+Qualitative Analysis of Prompt Setting To further evaluate the effect of dynamic demonstration and ROASP in SODP, we perform a human evaluation about the generated plannings (see Table 4) and show the planning results (see Fig. 6). For human evaluation, we randomly selected 100 REVERIE tasks and generate fine-grained step-by-step instructions in setting of SODP, SODP without dynamic demonstration, and SODP without ROASP. We asked 10 volunteers to mark the generated step-by-step instructions in terms of their relevancy and rationality. The relevancy score ranges from 0 (unrelated) to 3 (very related), which takes into account whether the keywords in instructions are related to the REVERIE task. For example, regarding the REVERIE instruction "Go to the kitchen and turn on the microwave", whether there are keywords in instructions related to the kitchen scene could be rated. Rationality is rated from 0 (bad) to 3 (perfect), considering whether the instruction conforms to the logic of navigation.
+
+The results are presented in Table 4. It shows that our SODP scored 2.06 on Relevancy and 1.93 on Rational-
+
+ity, which could be considered acceptable since the highest score is 3 and it is challenging to generate instructions that are consistent with tasks and actual navigation logic. When removing the dynamic demonstration, the score of generated instruction drops about $31.55\%$ on Relevancy and $36.27\%$ on Rationality, which could also be observed in Fig. 6. Although the instruction generated without dynamic demonstration is related to the task to some extent (e.g., "put the towel in the bottom drawer of the toilet" has the keyword "towel", the instruction lacks navigation information, such as how to reach the bathroom.) As shown in the bottom example of Fig. 6, instruction without ROASP successfully guided how to go to the destination location kitchen, but it still caused confusion by going upstairs and going downstairs several times, and thus reducing the rationality score, i.e. 1.64 on Relevancy and 1.55 on Rationality. More generation results can be found in the supplementary.
+
+# 6. Conclusions
+
+In this work, we propose a novel model, March-in-Chat (MiC), for the REVEIRE task, which only provides concise high-level instructions for the VLN agent. MiC enables the REVERIE agent to talk with an LLM on the fly to generate plans for the next few steps. It consists of three main modules, Goal-Oriented Static Planning (GOSP), Scene-Oriented Dynamic Planning (SODP), and Room-and-Object Aware Scene Perceiver (ROASP) module. We conduct extensive quantitative and qualitative experiments on REVERIE and the promising results show the effectiveness of our method.
+
+# 7. Acknowledgements
+
+Jing Liu is supported by the National Key Research and Development Program of China (No. 2020AAA0106400).
+
+# References
+
+[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Daniel Ho, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Eric Jang, Rosario Jauregui Ruano, Kyle Jeffrey, Sally Jesmonth, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Kuang-Huei Lee, Sergey Levine, Yao Lu, Linda Luu, Carolina Parada, Peter Pastor, Jornell Quiambao, Kanishka Rao, Jarek Rettinghouse, Diego Reyes, Pierre Sermanet, Nicolas Sievers, Clayton Tan, Alexander Toshev, Vincent Vanhoucke, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Mengyuan Yan, and Andy Zeng. Do as i can and not as i say: Grounding language in robotic affordances. In arXiv preprint arXiv:2204.01691, 2022. 3
+[2] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Topo-metric map pre-training for language-guided navigation. arXiv preprint arXiv:2212.04385, 2022. 2
+[3] Peter Anderson, Angel X. Chang, Devendra Singh Chaplot, Alexey Dosovitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, and Amir Roshan Zamir. On evaluation of embodied navigation agents. CoRR, abs/1807.06757, 2018. 6
+[4] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, pages 3674-3683, 2018. 1, 2, 5
+[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 3, 4
+[6] Angel X. Chang, Angela Dai, Thomas A. Funkhouser, Macej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from RGB-D data in indoor environments. In 3DV, pages 667-676, 2017. 5
+[7] Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. TOUCHDOWN: natural language navigation and spatial reasoning in visual street environments. In CVPR, pages 12538-12547, 2019. 1, 2
+[8] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021. 3
+[9] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, pages 5834-5847, 2021. 7
+
+[10] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Learning from unlabeled 3d environments for vision-and-language navigation. In ECCV, 2022. 4, 6, 7
+[11] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, 2022. 2, 7
+[12] Chen Gao, Jinyu Chen, Si Liu, Luting Wang, Qiong Zhang, and Qi Wu. Room-and-object aware knowledge reasoning for remote embodied referring expression. In CVPR, pages 3064-3073, 2021. 7
+[13] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In ICCV, pages 1634-1643, 2021. 7
+[14] Keji He, Yan Huang, Qi Wu, Jianhua Yang, Dong An, Shuanglin Sima, and Liang Wang. Landmark-rxr: Solving vision-and-language navigation with fine-grained alignment supervision. NeurIPS, 34:652-663, 2021. 2
+[15] Yicong Hong, Cristian Rodriguez Opazo, Qi Wu, and Stephen Gould. Sub-instruction aware vision-and-language navigation. In Bonnie Webber, Trevor Cohn, Yulan He, and Yang Liu, editors, EMNLP, pages 3360-3376, 2020. 5
+[16] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez Opazo, and Stephen Gould. VLN $\text{©}$ BERT: A recurrent vision-and-language BERT for navigation. In CVPR, pages 1643-1653, 2021. 2, 7
+[17] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In ICML, pages 9118-9147. PMLR, 2022. 2, 3, 5
+[18] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, Pierre Sermanet, Noah Brown, Tomas Jackson, Linda Luu, Sergey Levine, Karol Hausman, and Brian Ichter. Inner monologue: Embodied reasoning through planning with language models. In arXiv preprint arXiv:2207.05608, 2022. 3
+[19] Vihan Jain, Gabriel Magalhães, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. pages 1862-1872, 2019. 2
+[20] Melvin Johnson, Mike Schuster, Quoc V. Le, Maxim Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Thorat, Fernanda B. Viégas, Martin Wattenberg, Greg Corrado, Macduff Hughes, and Jeffrey Dean. Google's multilingual neural machine translation system: Enabling zero-shot translation. Trans. Assoc. Comput. Linguistics, 5:339-351, 2017. 6
+[21] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In EMNLP, pages 4392-4412, 2020. 1, 2
+[22] Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. What makes good in-context examples for gpt-3? arXiv preprint arXiv:2101.06804, 2021. 5
+
+[23] Chih-Yao Ma, Jiasen Lu, Zuxuan Wu, Ghassan AlRegib, Zsolt Kira, Richard Socher, and Caiming Xiong. Self-monitoring navigation agent via auxiliary progress estimation. In ICLR, 2019. 7
+[24] Khanh Nguyen and Hal Daumé III. Help, anna! visual navigation with natural multimodal assistance via retrospective curiosity-encouraging imitation learning. pages 684–695. Association for Computational Linguistics, 2019. 2
+[25] Gabriel Poesia, Alex Polozov, Vu Le, Ashish Tiwari, Gustavo Soares, Christopher Meek, and Sumit Gulwani. Synchromesh: Reliable code generation from pre-trained language models. In ICLR, 2022. 5
+[26] Yuankai Qi, Zizheng Pan, Yicong Hong, Ming-Hsuan Yang, Anton van den Hengel, and Qi Wu. The road to knowwhere: An object-and-room informed sequential bert for indoor vision-language navigation. In ICCV, pages 1655-1664, 2021. 2, 7
+[27] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. REVERIE: remote embodied visual referring expression in real indoor environments. In CVPR, pages 9979-9988, 2020. 1, 2, 6, 7
+[28] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. HOP: history-and-order aware pretraining for vision-and-language navigation. In CVPR, pages 15397-15406, 2022. 2, 7
+[29] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. Hop+: History-enhanced and order-aware pre-training for vision-and-language navigation. IEEE TPAMI, 2023. 2
+[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, volume 139, pages 8748-8763, 2021. 5
+[31] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3, 6
+[32] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors, EMNLP-IJCNLP, pages 3980-3990, 2019. 5
+[33] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A benchmark for interpreting grounded instructions for everyday tasks. In CVPR, pages 10737-10746, 2020. 1, 3
+[34] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. ProgPrompt: Generating situated robot task plans using large language models. 2022. 2
+[35] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. arXiv preprint arXiv:2212.04088, 2022. 3, 5
+
+[36] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In CoRL, pages 394-406, 2019. 1, 2
+[37] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In CVPR, pages 6629-6638, 2019. 7
+[38] Yusheng Zhao, Jinyu Chen, Chen Gao, Wenguan Wang, Lirong Yang, Haibing Ren, Huaxia Xia, and Si Liu. Target-driven structured transformer planner for vision-language navigation. pages 4194-4203, 2022. 7
+[39] Fengda Zhu, Xiwen Liang, Yi Zhu, Qizhi Yu, Xiaojun Chang, and Xiaodan Liang. SOON: scenario oriented object navigation with graph-based exploration. In CVPR, pages 12689-12699, 2021. 1, 2
\ No newline at end of file
diff --git a/2023/March in Chat_ Interactive Prompting for Remote Embodied Referring Expression/images.zip b/2023/March in Chat_ Interactive Prompting for Remote Embodied Referring Expression/images.zip
new file mode 100644
index 0000000000000000000000000000000000000000..34cc859d908b34659f70f759cc59cb879cf520d3
--- /dev/null
+++ b/2023/March in Chat_ Interactive Prompting for Remote Embodied Referring Expression/images.zip
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1c71e3333edc3167f5fd7306b32d09b7b5f97d1e368c547d55723037bdf0c30
+size 373820
diff --git a/2023/March in Chat_ Interactive Prompting for Remote Embodied Referring Expression/layout.json b/2023/March in Chat_ Interactive Prompting for Remote Embodied Referring Expression/layout.json
new file mode 100644
index 0000000000000000000000000000000000000000..98464b68a4f1083f96ebdfcb1e2abd7120910887
--- /dev/null
+++ b/2023/March in Chat_ Interactive Prompting for Remote Embodied Referring Expression/layout.json
@@ -0,0 +1,8676 @@
+{
+ "pdf_info": [
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 162,
+ 136,
+ 416,
+ 170
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 162,
+ 136,
+ 416,
+ 170
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 136,
+ 416,
+ 170
+ ],
+ "type": "text",
+ "content": "March in Chat: Interactive Prompting for Remote Embodied Referring Expression"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 126,
+ 191,
+ 454,
+ 207
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 126,
+ 191,
+ 454,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 191,
+ 454,
+ 207
+ ],
+ "type": "text",
+ "content": "Yanyuan Qiao"
+ },
+ {
+ "bbox": [
+ 126,
+ 191,
+ 454,
+ 207
+ ],
+ "type": "inline_equation",
+ "content": "^{1}"
+ },
+ {
+ "bbox": [
+ 126,
+ 191,
+ 454,
+ 207
+ ],
+ "type": "text",
+ "content": " Yuankai Qi"
+ },
+ {
+ "bbox": [
+ 126,
+ 191,
+ 454,
+ 207
+ ],
+ "type": "inline_equation",
+ "content": "^{1}"
+ },
+ {
+ "bbox": [
+ 126,
+ 191,
+ 454,
+ 207
+ ],
+ "type": "text",
+ "content": " Zheng Yu"
+ },
+ {
+ "bbox": [
+ 126,
+ 191,
+ 454,
+ 207
+ ],
+ "type": "inline_equation",
+ "content": "^{1}"
+ },
+ {
+ "bbox": [
+ 126,
+ 191,
+ 454,
+ 207
+ ],
+ "type": "text",
+ "content": " Jing Liu"
+ },
+ {
+ "bbox": [
+ 126,
+ 191,
+ 454,
+ 207
+ ],
+ "type": "inline_equation",
+ "content": "^{2,3}"
+ },
+ {
+ "bbox": [
+ 126,
+ 191,
+ 454,
+ 207
+ ],
+ "type": "text",
+ "content": " Qi Wu"
+ },
+ {
+ "bbox": [
+ 126,
+ 191,
+ 454,
+ 207
+ ],
+ "type": "inline_equation",
+ "content": "^{1*}"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 123,
+ 207,
+ 455,
+ 220
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 123,
+ 207,
+ 455,
+ 220
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 123,
+ 207,
+ 455,
+ 220
+ ],
+ "type": "inline_equation",
+ "content": "^{1}"
+ },
+ {
+ "bbox": [
+ 123,
+ 207,
+ 455,
+ 220
+ ],
+ "type": "text",
+ "content": "Australian Institute for Machine Learning, The University of Adelaide"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 157,
+ 220,
+ 420,
+ 233
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 157,
+ 220,
+ 420,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 157,
+ 220,
+ 420,
+ 233
+ ],
+ "type": "inline_equation",
+ "content": "^{2}"
+ },
+ {
+ "bbox": [
+ 157,
+ 220,
+ 420,
+ 233
+ ],
+ "type": "text",
+ "content": "Institute of Automation, Chinese Academy of Sciences"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 110,
+ 233,
+ 475,
+ 247
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 233,
+ 475,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 233,
+ 475,
+ 247
+ ],
+ "type": "inline_equation",
+ "content": "^{3}"
+ },
+ {
+ "bbox": [
+ 110,
+ 233,
+ 475,
+ 247
+ ],
+ "type": "text",
+ "content": "School of Artificial Intelligence, University of Chinese Academy of Sciences"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 68,
+ 249,
+ 508,
+ 261
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 68,
+ 249,
+ 508,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 68,
+ 249,
+ 508,
+ 261
+ ],
+ "type": "text",
+ "content": "{yanyuan.qiao,zheng.yu,qi.wu01}@adelaide.edu.au, qykshr@gmail.com,jliu@nlpr.ia.ac.cn"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 139,
+ 287,
+ 186,
+ 299
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 139,
+ 287,
+ 186,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 139,
+ 287,
+ 186,
+ 299
+ ],
+ "type": "text",
+ "content": "Abstract"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 45,
+ 313,
+ 282,
+ 581
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 313,
+ 282,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 313,
+ 282,
+ 581
+ ],
+ "type": "text",
+ "content": "Many Vision-and-Language Navigation (VLN) tasks have been proposed in recent years, from room-based to object-based and indoor to outdoor. The REVERIE (Remote Embodied Referring Expression) is interesting since it only provides high-level instructions to the agent, which are closer to human commands in practice. Nevertheless, this poses more challenges than other VLN tasks since it requires agents to infer a navigation plan only based on a short instruction. Large Language Models (LLMs) show great potential in robot action planning by providing proper prompts. Still, this strategy has not been explored under the REVERIE settings. There are several new challenges. For example, the LLM should be environment-aware so that the navigation plan can be adjusted based on the current visual observation. Moreover, the LLM planned actions should be adaptable to the much larger and more complex REVERIE environment. This paper proposes a March-in-Chat (MiC) model that can talk to the LLM on the fly and plan dynamically based on a newly proposed Room-and-Object Aware Scene Perceiver (ROASP). Our MiC model outperforms the previous state-of-the-art by large margins by SPL and RGSPL metrics on the REVERIE benchmark. The source code is available at https://github.com/YanyuanQiao/MiC"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 46,
+ 606,
+ 124,
+ 618
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 606,
+ 124,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 606,
+ 124,
+ 618
+ ],
+ "type": "text",
+ "content": "1. Introduction"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 45,
+ 626,
+ 279,
+ 708
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 626,
+ 279,
+ 708
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 626,
+ 279,
+ 708
+ ],
+ "type": "text",
+ "content": "Vision-and-Language Navigation (VLN), which lies at the intersection of computer vision, natural language processing and robotics, has aroused great attention from research communities in the past few years. Given instructions in natural language, the VLN agent should navigate to the target location based on the dynamic observations in the 3D simulated environments. Since VLN has great"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 309,
+ 312,
+ 516,
+ 433
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 305,
+ 291,
+ 516,
+ 310
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 305,
+ 291,
+ 516,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 305,
+ 291,
+ 516,
+ 310
+ ],
+ "type": "text",
+ "content": "REVERIE Instruction: Clean the small picture in front of the large mirror."
+ }
+ ]
+ }
+ ],
+ "index": 12,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 309,
+ 312,
+ 516,
+ 433
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 309,
+ 312,
+ 516,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 309,
+ 312,
+ 516,
+ 433
+ ],
+ "type": "image",
+ "image_path": "8f6ac22cb1ca7ce66006da99c22e278d7607c56e42a86de54ec831bc88a4072c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 296,
+ 442,
+ 531,
+ 490
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 296,
+ 442,
+ 531,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 296,
+ 442,
+ 531,
+ 490
+ ],
+ "type": "text",
+ "content": "Figure 1: Our March-in-Chat (MiC) model is talking to a Large Language Model (LLM) to generate navigation plans on the fly, with the REVERIE instruction and the dynamic room-and-object information as inputs."
+ }
+ ]
+ }
+ ],
+ "index": 14,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 296,
+ 496,
+ 530,
+ 624
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 296,
+ 496,
+ 530,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 296,
+ 496,
+ 530,
+ 624
+ ],
+ "type": "text",
+ "content": "potential in real-world applications such as domestic assistant robots, a large amount of specific VLN tasks have been proposed, including R2R [4] and RxR [21] that ask the agent to navigate from one room to another in a photorealistic environment according to the fine-grained instruction, NDH [36] provides detailed dialogues which imply the instruction, TouchDown [7] extends the task into an outdoor environment, REVERIE [27] and SOON [39] that additionally require the agent's ability of remote object grounding and ALFRED [33] that asks the agent to interact with the target object in a single room of the synthetic environment."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 296,
+ 624,
+ 531,
+ 730
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 296,
+ 624,
+ 531,
+ 730
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 296,
+ 624,
+ 531,
+ 730
+ ],
+ "type": "text",
+ "content": "Most of these VLN tasks provide detailed step-by-step instructions to the agent, such as \"Go up the stairs and then walk the length of the couch. Walk past the dining area and into the kitchen. Stop in front of the refrigerator.\" in R2R. Although detailed instructions can help the agent better achieve the navigation goal in the simulated environments, it has a big gap towards real applications where human beings tend to give coarse-grained high-level instructions such as \"Go to the refrigerator on the second floor\"."
+ }
+ ]
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 77,
+ 42,
+ 119,
+ 72
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 42,
+ 119,
+ 72
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 42,
+ 119,
+ 72
+ ],
+ "type": "text",
+ "content": "CVF"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 151,
+ 41,
+ 470,
+ 73
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 151,
+ 41,
+ 470,
+ 73
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 151,
+ 41,
+ 470,
+ 73
+ ],
+ "type": "text",
+ "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 57,
+ 719,
+ 131,
+ 730
+ ],
+ "type": "page_footnote",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 57,
+ 719,
+ 131,
+ 730
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 57,
+ 719,
+ 131,
+ 730
+ ],
+ "type": "text",
+ "content": "*Corresponding author"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 277,
+ 746,
+ 301,
+ 756
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 277,
+ 746,
+ 301,
+ 756
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 277,
+ 746,
+ 301,
+ 756
+ ],
+ "type": "text",
+ "content": "15758"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "page_size": [
+ 595,
+ 842
+ ],
+ "page_idx": 0
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 45,
+ 106,
+ 279,
+ 212
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 106,
+ 279,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 106,
+ 279,
+ 212
+ ],
+ "type": "text",
+ "content": "Contrary to other tasks, the Remote Embodied Referring Expression (REVERIE) task is more likely to empower the real-world applications of VLN, of which the instructions are closer to those in practice, such as \"Empty the washing machine on level one\". Such high-level instruction is more challenging for VLN agents since it requires them to be more competent in perceiving the surrounding environment and the navigation progress and correspondingly making reasonable plans for the next steps."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 45,
+ 213,
+ 280,
+ 341
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 213,
+ 280,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 213,
+ 280,
+ 341
+ ],
+ "type": "text",
+ "content": "Recently, Large Language Models (LLMs) that internalize a wealth of commonsense knowledge show great potential in action planning for some embodied tasks with the help of suitable in-context learning. However, previous works mainly utilize LLMs to plan atomic actions of object manipulation in a very limited space with simple scenes. These predefined atomic actions can be easily planned well by the LLMs planners with a unified template. Different from these embodied tasks, REVERIE requires large-area exploration from one room to another, which is complex in the layout of rooms and scenes with diverse objects."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 45,
+ 343,
+ 280,
+ 588
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 343,
+ 280,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 343,
+ 280,
+ 588
+ ],
+ "type": "text",
+ "content": "In this work, to adapt LLMs as the planner for REVERIE with the ability of comprehensive scene perception, we propose a novel model named March in Chat (MiC), which enables the LLM as an environment-aware instruction planner through on-the-fly dialogues between the agent and the LLM as Fig. 1 shows. Specifically, the agent is initially situated at the starting position given a high-level coarse-grained REVERIE instruction. First, a Goal-Oriented Static Planning (GOSP) module queries the LLM to point out the target object and infer where the thing may be by using the rich world knowledge internalized in the LLM. Secondly, the agent's Room-and-Object Aware Scene Perceiver (ROASP) describes the current observation and asks the LLM to generate step-by-step fine-grained planning for the next navigation steps. Then, if the ROASP finds the room has changed, the LLM is queried again by the Scene-Oriented Dynamic Planning (SODP) module to generate a new fine-grained step-by-step planning, which will be concatenated with all previous responses from the LLM. The agent will march under the guidance of such interactive prompting until the task is finished."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 45,
+ 590,
+ 279,
+ 730
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 590,
+ 279,
+ 730
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 590,
+ 279,
+ 730
+ ],
+ "type": "text",
+ "content": "To evaluate our proposed MiC, we conduct experiments on the REVERIE benchmark. Our MiC achieves a new state-of-the-art performance in all metrics on REVERIE val unseen set and REVERIE test unseen set. Mainly, MiC obtains "
+ },
+ {
+ "bbox": [
+ 45,
+ 590,
+ 279,
+ 730
+ ],
+ "type": "inline_equation",
+ "content": "41.97\\%"
+ },
+ {
+ "bbox": [
+ 45,
+ 590,
+ 279,
+ 730
+ ],
+ "type": "text",
+ "content": " on the primary navigation metric of SPL and "
+ },
+ {
+ "bbox": [
+ 45,
+ 590,
+ 279,
+ 730
+ ],
+ "type": "inline_equation",
+ "content": "26.17\\%"
+ },
+ {
+ "bbox": [
+ 45,
+ 590,
+ 279,
+ 730
+ ],
+ "type": "text",
+ "content": " on the major object grounding metric of RGSPL on test split, which is at least "
+ },
+ {
+ "bbox": [
+ 45,
+ 590,
+ 279,
+ 730
+ ],
+ "type": "inline_equation",
+ "content": "3.09\\%"
+ },
+ {
+ "bbox": [
+ 45,
+ 590,
+ 279,
+ 730
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 45,
+ 590,
+ 279,
+ 730
+ ],
+ "type": "inline_equation",
+ "content": "3.49\\%"
+ },
+ {
+ "bbox": [
+ 45,
+ 590,
+ 279,
+ 730
+ ],
+ "type": "text",
+ "content": " higher than the previous SoTA results. We also conduct ablation studies to validate the contributions of different components in MiC and the effect of scene-aware perception in dynamic planning generation. These promising results demonstrate the effectiveness of our proposed MiC."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 308,
+ 106,
+ 508,
+ 117
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 106,
+ 508,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 106,
+ 508,
+ 117
+ ],
+ "type": "text",
+ "content": "In summary, we make the following contributions:"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 308,
+ 117,
+ 530,
+ 246
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 8,
+ "blocks": [
+ {
+ "bbox": [
+ 308,
+ 117,
+ 530,
+ 152
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 117,
+ 530,
+ 152
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 117,
+ 530,
+ 152
+ ],
+ "type": "text",
+ "content": "- We propose a novel March-in-Chat (MiC) model, which lets the REVERIE agent talk with an LLM on the fly to make plans for the next few steps."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 308,
+ 154,
+ 530,
+ 212
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 154,
+ 530,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 154,
+ 530,
+ 212
+ ],
+ "type": "text",
+ "content": "- Two planning modules, namely Goal-Oriented Static Planning (GOSP) module, and Scene-Oriented Dynamic Planning (SODP) module, and one Room-and-Object Aware Scene Perceiver (ROASP) module, are proposed."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 308,
+ 213,
+ 530,
+ 246
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 213,
+ 530,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 213,
+ 530,
+ 246
+ ],
+ "type": "text",
+ "content": "- Extensive quantitative and qualitative experiments are conducted on REVERIE to validate the effectiveness of our method."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 297,
+ 256,
+ 379,
+ 268
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 297,
+ 256,
+ 379,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 297,
+ 256,
+ 379,
+ 268
+ ],
+ "type": "text",
+ "content": "2. Related work"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 296,
+ 276,
+ 530,
+ 554
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 296,
+ 276,
+ 530,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 296,
+ 276,
+ 530,
+ 554
+ ],
+ "type": "text",
+ "content": "Vision-and-Language Navigation Vision-and-Language Navigation (VLN) has attracted increasing attention in recent years, and many specific VLN tasks have been proposed [7, 19, 24, 27, 29, 36]. Anderson et al. [4] proposes the first VLN benchmark, Room-to-Room (R2R), which requires an agent to navigate from one room to another in a house, according to a detailed natural language step-by-step instruction in a photo-realistic environment. Later, Room-across-Room (RxR) [21] was proposed with longer and more detailed multilingual instructions. Both these two tasks give fine-grained instructions, which makes it easier to navigate to the target location. NDH [36] extends the navigation instruction to the dialogue form, and TouchDown [7] extends the environments to outdoor. REVERIE [27] and SOON [39] are proposed for remote object localization, which requires an agent not only to navigate to the target location, but also to specify the object to interact with. The difference between REVERIE and SOON is that REVERIE uses short concise instructions (e.g., \"bring me the red cup from the kitchen.\") while SOON employs long detailed instructions (e.g., \"I want to find a cylindrical, metallic and tall lamp, which is set in the bright living room. The lamp is on the cabinet which is on the left of the television and next to the window. The living room is on ...\")."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 296,
+ 555,
+ 530,
+ 706
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 296,
+ 555,
+ 530,
+ 706
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 296,
+ 555,
+ 530,
+ 706
+ ],
+ "type": "text",
+ "content": "Among the aforementioned VLN tasks, the instructions of REVERIE are closer to what we would say to an intelligent domestic robot in daily life in terms of the instruction length and logic, which is usually short and concise. However, most existing methods [2, 11, 14, 16, 26, 28] are usually designed for the VLN tasks where detailed step-by-step instructions are used, thus they do not perform well on REVERIE. In this work, we specifically develop a method for REVERIE. Inspired by the fact of LLMs that implicitly internalize rich knowledge in action planning, we propose to exploit LLM as a fine-grained planner to generate detailed navigation plans from the concise instructions of REVERIE to improve navigation success."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 297,
+ 706,
+ 530,
+ 730
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 297,
+ 706,
+ 530,
+ 730
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 297,
+ 706,
+ 530,
+ 730
+ ],
+ "type": "text",
+ "content": "LLMs as Embodied Planner Benefiting from the rise of LLMs, recent works [17, 34] have explored the use of LLMs"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 277,
+ 746,
+ 301,
+ 756
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 277,
+ 746,
+ 301,
+ 756
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 277,
+ 746,
+ 301,
+ 756
+ ],
+ "type": "text",
+ "content": "15759"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "page_size": [
+ 595,
+ 842
+ ],
+ "page_idx": 1
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 311,
+ 117,
+ 320,
+ 125
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 311,
+ 117,
+ 320,
+ 125
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 311,
+ 117,
+ 320,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 311,
+ 117,
+ 320,
+ 125
+ ],
+ "type": "image",
+ "image_path": "2d287ec19b9b3dacf1b36e1f271e5271bca29e9fc29f2915cae0c7ef9fcdc46a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 349,
+ 117,
+ 367,
+ 125
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 349,
+ 117,
+ 367,
+ 125
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 349,
+ 117,
+ 367,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 349,
+ 117,
+ 367,
+ 125
+ ],
+ "type": "image",
+ "image_path": "2da1df1ad04dc53bc1e070729fc1fa91427a6ca4d5af948960bc2bbf0b0e3568.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 468,
+ 117,
+ 477,
+ 125
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 468,
+ 117,
+ 477,
+ 125
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 468,
+ 117,
+ 477,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 468,
+ 117,
+ 477,
+ 125
+ ],
+ "type": "image",
+ "image_path": "1b049502b19f12fe0ae918c8b89a5c491aa4c61228cb763ae4f1c5175ffce99e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 54,
+ 127,
+ 520,
+ 268
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 54,
+ 113,
+ 277,
+ 124
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 54,
+ 113,
+ 277,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 54,
+ 113,
+ 277,
+ 124
+ ],
+ "type": "text",
+ "content": "REVERIE instruction: Empty the washing machine on level one."
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 54,
+ 127,
+ 520,
+ 268
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 54,
+ 127,
+ 520,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 54,
+ 127,
+ 520,
+ 268
+ ],
+ "type": "image",
+ "image_path": "ddb2fd4827834db8178e14901a821a6aa28659560df8d09f7fc9cf04195e11b2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 45,
+ 277,
+ 531,
+ 348
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 277,
+ 531,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 277,
+ 531,
+ 348
+ ],
+ "type": "text",
+ "content": "Figure 2: Overview of our March-in-Chat model. The program runs along the vertical arrows from left to right, progressing with the time flow. Our model first performs Goal-Oriented Static Planning (GOSP, Sec. 4.1.1) to reason the target object and its possible lying room; then the Room-and-Object Aware Scene Perceiver (ROASP, Sec. 4.2) perceives what room type the agent currently stands in and what prominent objects can be seen; these information are used by the Scene-Oriented Dynamic Planning module (SODP, Sec. 4.1.2) to generate a detailed instruction to execute. GOSP just runs once, and we repeat ROASP and SODP until the agent chooses to stop or reaches the maximum steps."
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 45,
+ 358,
+ 279,
+ 521
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 358,
+ 279,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 358,
+ 279,
+ 521
+ ],
+ "type": "text",
+ "content": "in task planning for various embodied tasks. Huang et al. [17] propose to utilize the frozen LLMs (e.g., GPT-2 [31], GPT-3 [5] and Codex [8]) to plan actions for the embodied agent with in-context learning [5]. SayCan [1] translates a high-level instruction into a list of candidate low-level actions with a probability, which is then multiplied by a value function for action prediction. These two LLM planners are static, which only generate action plans at the beginning of a task. By contrast, Huang et al. [18] propose to introduce the feedback of action progress, detected objects and human assistance into the LLM planner to re-plan atomic actions. One concurrent work by Song et al. [35] injects the detected objects to re-generate high-level plans with a fixed program pattern for the ALFRED [33] task."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 45,
+ 522,
+ 279,
+ 639
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 522,
+ 279,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 522,
+ 279,
+ 639
+ ],
+ "type": "text",
+ "content": "However, these above-mentioned methods mainly concentrate on planning atomic actions for object manipulation in a very limited space with simple scenes. By contrast, REVERIE has plenty of much larger and more complicated environments: 90 multi-layer buildings of various styles (e.g., office, home, gym, to name a few). To handle the complex scenarios of REVERIE, we propose a Room-and-Object Aware Scene Perceiver module that helps the LLM planner dynamically interact with the environment in the form of a natural language dialogue."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 46,
+ 650,
+ 135,
+ 664
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 650,
+ 135,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 650,
+ 135,
+ 664
+ ],
+ "type": "text",
+ "content": "3. Problem Setup"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 45,
+ 671,
+ 279,
+ 730
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 671,
+ 279,
+ 730
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 671,
+ 279,
+ 730
+ ],
+ "type": "text",
+ "content": "In the REVERIE task, given a concise and high-level instruction referring to a remote object, the agent is expected to navigate to the goal location and identify the target object in previously unseen environments. The environment is defined as an undirected graph "
+ },
+ {
+ "bbox": [
+ 45,
+ 671,
+ 279,
+ 730
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}"
+ },
+ {
+ "bbox": [
+ 45,
+ 671,
+ 279,
+ 730
+ ],
+ "type": "text",
+ "content": ", where"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{V} = \\{V_i\\}_{i=1}^K"
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "text",
+ "content": " denotes "
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "inline_equation",
+ "content": "K"
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "text",
+ "content": " navigable nodes, and "
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{E}"
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "text",
+ "content": " denotes connectivity edges. The agent is first placed in a starting node with the initial state "
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "inline_equation",
+ "content": "s_0"
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "text",
+ "content": " and perceives a panorama "
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{R}_t"
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "text",
+ "content": " as the visual observation at each time step "
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "inline_equation",
+ "content": "t"
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "text",
+ "content": ". The panorama "
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{R}_t"
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "text",
+ "content": " is split into "
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "inline_equation",
+ "content": "n"
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "text",
+ "content": " single view images as "
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{R}_t = \\{r_i\\}_{i=1}^n"
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "text",
+ "content": ". Each single view image "
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "inline_equation",
+ "content": "r_i"
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "text",
+ "content": " is represented by an image feature vector and an orientation feature vector. In addition, the object features "
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{O}_t = \\{o_i\\}_{i=1}^m"
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "text",
+ "content": " of "
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "inline_equation",
+ "content": "m"
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "text",
+ "content": " objects are extracted from the panorama view using the annotated object bounding boxes or object detectors. Then, the agent makes a sequence of actions "
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "inline_equation",
+ "content": "\\langle a_0, \\ldots, a_T \\rangle"
+ },
+ {
+ "bbox": [
+ 296,
+ 357,
+ 530,
+ 533
+ ],
+ "type": "text",
+ "content": " to reach the target location, where each action is achieved by choosing a navigable node from the candidate list. The agent navigates in the environment until the target object is grounded or the agent reaches the predefined maximum trajectory length."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 297,
+ 545,
+ 351,
+ 556
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 297,
+ 545,
+ 351,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 297,
+ 545,
+ 351,
+ 556
+ ],
+ "type": "text",
+ "content": "4. Method"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 296,
+ 566,
+ 530,
+ 730
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 296,
+ 566,
+ 530,
+ 730
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 296,
+ 566,
+ 530,
+ 730
+ ],
+ "type": "text",
+ "content": "As illustrated in Fig. 2, when initially situated at the starting position and given a concise high-level instruction such as \"Empty the washing machine on level one\", the agent first queries an LLM with the GOSP module (Sec. 4.1.1) to find out the target object \"washing machine\" in the instruction and reason out the potential location \"laundry room\" by using the world knowledge implied in an LLM. Then the ROASP module (Sec. 4.2) extracts the room type and visible objects from the current visual observation to obtain the environmental feedback. With the description of the scene perception, the LLM is queried again by the SODP (Sec. 4.1.2) to generate the next step instruction, which is used to guide the agent to navigate to the target object in the room. The baseline agent is based on"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 277,
+ 746,
+ 301,
+ 756
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 277,
+ 746,
+ 301,
+ 756
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 277,
+ 746,
+ 301,
+ 756
+ ],
+ "type": "text",
+ "content": "15760"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "page_size": [
+ 595,
+ 842
+ ],
+ "page_idx": 2
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 133,
+ 113,
+ 227,
+ 123
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 113,
+ 227,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 113,
+ 227,
+ 123
+ ],
+ "type": "text",
+ "content": "Target Object Recognition"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 90,
+ 124,
+ 113,
+ 132
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 90,
+ 124,
+ 113,
+ 132
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 90,
+ 124,
+ 113,
+ 132
+ ],
+ "type": "text",
+ "content": "Task:"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 90,
+ 133,
+ 254,
+ 142
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 90,
+ 133,
+ 254,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 90,
+ 133,
+ 254,
+ 142
+ ],
+ "type": "text",
+ "content": "Empty the washing machine on level one."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 89,
+ 147,
+ 113,
+ 154
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 89,
+ 147,
+ 113,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 89,
+ 147,
+ 113,
+ 154
+ ],
+ "type": "text",
+ "content": "Goal:"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 90,
+ 156,
+ 247,
+ 165
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 90,
+ 156,
+ 247,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 90,
+ 156,
+ 247,
+ 165
+ ],
+ "type": "text",
+ "content": "The target object is: washing machine"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 133,
+ 178,
+ 229,
+ 187
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 178,
+ 229,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 178,
+ 229,
+ 187
+ ],
+ "type": "text",
+ "content": "Target Object Localization"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 87,
+ 187,
+ 122,
+ 197
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 87,
+ 187,
+ 122,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 87,
+ 187,
+ 122,
+ 197
+ ],
+ "type": "text",
+ "content": "Example:"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 87,
+ 197,
+ 274,
+ 214
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 87,
+ 197,
+ 274,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 87,
+ 197,
+ 274,
+ 214
+ ],
+ "type": "text",
+ "content": "Question: Where does a microwave can usually appear in a house?"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 87,
+ 215,
+ 155,
+ 223
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 87,
+ 215,
+ 155,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 87,
+ 215,
+ 155,
+ 223
+ ],
+ "type": "text",
+ "content": "Answer: kitchen."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 87,
+ 229,
+ 266,
+ 247
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 87,
+ 229,
+ 266,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 87,
+ 229,
+ 266,
+ 247
+ ],
+ "type": "text",
+ "content": "Question: Where does a washing machine can usually appear in a house?"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 87,
+ 248,
+ 174,
+ 256
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 87,
+ 248,
+ 174,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 87,
+ 248,
+ 174,
+ 256
+ ],
+ "type": "text",
+ "content": "Answer: laundry room"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 98,
+ 262,
+ 264,
+ 271
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 98,
+ 262,
+ 264,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 98,
+ 262,
+ 264,
+ 271
+ ],
+ "type": "text",
+ "content": "(a) Prompt template of Goal-Oriented Static Planning"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 333,
+ 112,
+ 440,
+ 122
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 333,
+ 112,
+ 440,
+ 122
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 333,
+ 112,
+ 440,
+ 122
+ ],
+ "type": "text",
+ "content": "Scene-Oriented Dynamic Plan"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 291,
+ 132,
+ 482,
+ 151
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 291,
+ 132,
+ 482,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 291,
+ 132,
+ 482,
+ 151
+ ],
+ "type": "text",
+ "content": "At this step, I am in bedroom, I can see bed, lamp, pillow."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 292,
+ 159,
+ 326,
+ 168
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 292,
+ 159,
+ 326,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 292,
+ 159,
+ 326,
+ 168
+ ],
+ "type": "text",
+ "content": "Example:"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 292,
+ 169,
+ 445,
+ 178
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 292,
+ 169,
+ 445,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 292,
+ 169,
+ 445,
+ 178
+ ],
+ "type": "text",
+ "content": "Step 1: go straight and out the room"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 292,
+ 178,
+ 411,
+ 186
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 292,
+ 178,
+ 411,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 292,
+ 178,
+ 411,
+ 186
+ ],
+ "type": "text",
+ "content": "Step 2: pass the white table"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 292,
+ 187,
+ 482,
+ 205
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 292,
+ 187,
+ 482,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 292,
+ 187,
+ 482,
+ 205
+ ],
+ "type": "text",
+ "content": "Step 3: move into the laundry room across the hall"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 292,
+ 205,
+ 394,
+ 214
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 292,
+ 205,
+ 394,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 292,
+ 205,
+ 394,
+ 214
+ ],
+ "type": "text",
+ "content": "Step 4: stop at the sink"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 291,
+ 223,
+ 481,
+ 242
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 291,
+ 223,
+ 481,
+ 242
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 291,
+ 223,
+ 481,
+ 242
+ ],
+ "type": "text",
+ "content": "Task: Empty the washing machine on level one. Step 1: exit the bedroom"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 294,
+ 262,
+ 474,
+ 271
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 294,
+ 262,
+ 474,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 294,
+ 262,
+ 474,
+ 271
+ ],
+ "type": "text",
+ "content": "(b) Prompt template of Scene-Oriented Dynamic Planning"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 45,
+ 275,
+ 530,
+ 299
+ ],
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 275,
+ 530,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 275,
+ 530,
+ 299
+ ],
+ "type": "text",
+ "content": "Figure 3: Examples of prompting templates for Goal-Oriented Static Planning (a) and Scene-Oriented Dynamic Planning (b). Outputs are marked with green color. And the red denotes the predicted object from Target Object Recognition."
+ }
+ ]
+ }
+ ],
+ "index": 21,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 45,
+ 311,
+ 279,
+ 335
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 311,
+ 279,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 311,
+ 279,
+ 335
+ ],
+ "type": "text",
+ "content": "HM3D-DUET [10], more details can be found in the supplementary material."
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 45,
+ 336,
+ 272,
+ 348
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 336,
+ 272,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 336,
+ 272,
+ 348
+ ],
+ "type": "text",
+ "content": "4.1. Planning with World Knowledge from LLMs"
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 45,
+ 354,
+ 279,
+ 424
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 354,
+ 279,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 354,
+ 279,
+ 424
+ ],
+ "type": "text",
+ "content": "This section illustrates how we utilize in-context learning (ICL) [5] to acquire world knowledge from the LLMs for planning. We first briefly introduce in-context learning. Then, we elaborate the Goal-Oriented Static Planning (GOSP) and Scene-Oriented Dynamic Planning (SODP). Last, we show the demonstration selection process."
+ }
+ ]
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 45,
+ 425,
+ 279,
+ 552
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 425,
+ 279,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 425,
+ 279,
+ 552
+ ],
+ "type": "text",
+ "content": "Preliminary: In-context Learning for Planning. In-context learning (ICL) is a paradigm that lets LLMs directly make predictions based on a natural language context without gradient updates [5]. Specifically, under the setting of in-context learning, an LLM is fed a \"prompt\" that usually contains a task description and several demonstrations, and then the LLM generates the required outputs. Both the prompt template and the choice of demonstration examples have an impact on how well ICL performs. In this work, we use two different ICL settings to generate different navigation plans, i.e. the GOSP and SODP module."
+ }
+ ]
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 45,
+ 554,
+ 279,
+ 670
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 554,
+ 279,
+ 670
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 554,
+ 279,
+ 670
+ ],
+ "type": "text",
+ "content": "The GOSP aims to identify the target object and infer the target location by arousing the world knowledge contained in an LLM through appropriate prompts. A fixed demonstration example is used for GOSP. While the SODP aims to generate step-by-step planning instructions after observing the dynamic scenes from the environment, which is more complicated than the former. To better generate plannings, we dynamically select the most suitable demonstration examples for SODP and incorporate the environmental feedback as prompts for interactive planning."
+ }
+ ]
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 45,
+ 679,
+ 239,
+ 691
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 679,
+ 239,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 679,
+ 239,
+ 691
+ ],
+ "type": "text",
+ "content": "4.1.1 Goal-Oriented Static Planning (GOSP)"
+ }
+ ]
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 45,
+ 694,
+ 279,
+ 730
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 694,
+ 279,
+ 730
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 694,
+ 279,
+ 730
+ ],
+ "type": "text",
+ "content": "Given a high-level concise instruction, such as \"Empty the washing machine on level one\", an LLM is first asked to generate a goal-oriented static planning instruction: \"Goal:"
+ }
+ ]
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 296,
+ 311,
+ 530,
+ 591
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 296,
+ 311,
+ 530,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 296,
+ 311,
+ 530,
+ 591
+ ],
+ "type": "text",
+ "content": "The target object is a washing machine. It is usually in a laundry room\", which emphasizes the target object and points out where the target object may lie. As shown in Fig. 3(a), the planning generation mainly consists of two sub-tasks: target object recognition and target object localization, which can be achieved by providing specifically designed prompts for the LLM. To this end, we design the prompts for the former sub-task in the form: \"Task: Empty the washing machine on level one. Goal: The target object is:\". Then the LLM will generate a corresponding answer: \"washing machine\". By contrast, the latter subtask, reasoning out the target location, is more complex because it requires more suitable prompts to arouse the internalized world knowledge in the LLM. To address this problem, we utilize a fixed demonstration example for the LLM, and design the prompts for the latter sub-task in the form: \"Example: Question: Where does a microwave can usually appear in a house? Answer: kitchen. Question: Where does a washing machine can usually appear in a house? Answer: \". Given such prompts, the LLM will generate the corresponding answer \"laundry room\". With these answers, we can easily combine them into the goal-oriented planning format: \"Goal: The target object is a washing machine. It is usually in a laundry room\"."
+ }
+ ]
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 296,
+ 598,
+ 508,
+ 610
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 296,
+ 598,
+ 508,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 296,
+ 598,
+ 508,
+ 610
+ ],
+ "type": "text",
+ "content": "4.1.2 Scene-Oriented Dynamic Planning (SODP)"
+ }
+ ]
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 296,
+ 613,
+ 530,
+ 730
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 296,
+ 613,
+ 530,
+ 730
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 296,
+ 613,
+ 530,
+ 730
+ ],
+ "type": "text",
+ "content": "As is shown in Fig 3(b), the prompt for SODP consists of three parts. The first part is based on the scene perception of room type, such as \"bedroom\", and visible objects, such as \"bed, lamp, pillow\", obtained by the Room-and-Object Aware Scene Perceiver (ROASP, Sec. 4.2). These information are transformed into a natural language description of the current scene in the format of \"At this step, I am in bedroom, I can see bed, lamp, pillow\". The second part is a demonstration of the fine-grained step-by-step instruction, which is selected according to the strategy"
+ }
+ ]
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 277,
+ 747,
+ 299,
+ 756
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 277,
+ 747,
+ 299,
+ 756
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 277,
+ 747,
+ 299,
+ 756
+ ],
+ "type": "text",
+ "content": "15761"
+ }
+ ]
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "page_size": [
+ 595,
+ 842
+ ],
+ "page_idx": 3
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 54,
+ 109,
+ 70,
+ 117
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 54,
+ 109,
+ 70,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 54,
+ 109,
+ 70,
+ 117
+ ],
+ "type": "text",
+ "content": "R2R"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 54,
+ 117,
+ 263,
+ 134
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 54,
+ 117,
+ 263,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 54,
+ 117,
+ 263,
+ 134
+ ],
+ "type": "text",
+ "content": "Go up the stairs and then walk the length of the couch. Walk past the dining area and into the kitchen. Stop in front of the refrigerator."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 54,
+ 137,
+ 80,
+ 143
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 54,
+ 137,
+ 80,
+ 143
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 54,
+ 137,
+ 80,
+ 143
+ ],
+ "type": "text",
+ "content": "FGR2R"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 54,
+ 145,
+ 200,
+ 178
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 7,
+ "blocks": [
+ {
+ "bbox": [
+ 54,
+ 145,
+ 117,
+ 154
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 54,
+ 145,
+ 117,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 54,
+ 145,
+ 117,
+ 154
+ ],
+ "type": "text",
+ "content": "Step 1: go up the stair"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 54,
+ 154,
+ 179,
+ 161
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 54,
+ 154,
+ 179,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 54,
+ 154,
+ 179,
+ 161
+ ],
+ "type": "text",
+ "content": "Step 2: and then walk the length of the couch"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 54,
+ 162,
+ 200,
+ 170
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 54,
+ 162,
+ 200,
+ 170
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 54,
+ 162,
+ 200,
+ 170
+ ],
+ "type": "text",
+ "content": "Step 3: walk past the dining area and into the kitchen"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 54,
+ 170,
+ 160,
+ 178
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 54,
+ 170,
+ 160,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 54,
+ 170,
+ 160,
+ 178
+ ],
+ "type": "text",
+ "content": "Step 4: stop in front of the refrigerator"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 54,
+ 186,
+ 269,
+ 197
+ ],
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 54,
+ 186,
+ 269,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 54,
+ 186,
+ 269,
+ 197
+ ],
+ "type": "text",
+ "content": "Figure 4: Example of instructions in R2R and FGR2R."
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 45,
+ 208,
+ 279,
+ 278
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 208,
+ 279,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 208,
+ 279,
+ 278
+ ],
+ "type": "text",
+ "content": "detailed in the next section. The last part is previous instructions, such as \"Task: Empty the washing machine on level one. Step 1:\". All these three parts are concatenated together and then fed into an LLM to generate the fine-grained planning instruction for the next step accordingly, such as \"Exit the bedroom\"."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 45,
+ 286,
+ 219,
+ 297
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 286,
+ 219,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 286,
+ 219,
+ 297
+ ],
+ "type": "text",
+ "content": "4.1.3 Dynamic Demonstration Selection"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 45,
+ 300,
+ 279,
+ 393
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 300,
+ 279,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 300,
+ 279,
+ 393
+ ],
+ "type": "text",
+ "content": "Recent works show that providing various demonstration examples to LLMs benefits the in-context learning for different tasks [17, 22, 25]. In light of these findings, to direct the LLMs in generating better fine-grained plannings, we dynamically select the most suitable demonstration example for each specific task in REVERIE as the prompt to generate the environment-aware instruction, contrary to using a single fixed demonstration for all tasks."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "text",
+ "content": "Specifically, we choose the training set of the Fine-Grained R2R dataset (FGR2R) [15] as the demonstration set "
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{D}"
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "text",
+ "content": ", of which each sample will be used as a demonstration example "
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "inline_equation",
+ "content": "D_{step}"
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "text",
+ "content": ". As shown in Fig. 4, FGR2R decomposes each low-level instruction "
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "inline_equation",
+ "content": "I_{low}"
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "text",
+ "content": " of R2R dataset [4] into step-by-step instructions "
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "inline_equation",
+ "content": "I_{step}"
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "text",
+ "content": ". Then, given a high-level instruction "
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "inline_equation",
+ "content": "I_{high}"
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "text",
+ "content": " of REVERIE, a proper "
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "inline_equation",
+ "content": "I_{low}"
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "text",
+ "content": " will be selected as the demonstration example "
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "inline_equation",
+ "content": "D_{step}"
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "text",
+ "content": " by a matching algorithm. In particular, we use "
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "inline_equation",
+ "content": "I_{high}"
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "text",
+ "content": " as query "
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "inline_equation",
+ "content": "Q"
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "text",
+ "content": " and each low-level instruction "
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "inline_equation",
+ "content": "I_{low}"
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "text",
+ "content": " as the key "
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "inline_equation",
+ "content": "K_i"
+ },
+ {
+ "bbox": [
+ 45,
+ 394,
+ 279,
+ 545
+ ],
+ "type": "text",
+ "content": ", both of which are embedded by the Sentence-BERT [32]. The semantic distance score between the two embeddings is calculated by the cosine similarity:"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 104,
+ 554,
+ 279,
+ 580
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 554,
+ 279,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 554,
+ 279,
+ 580
+ ],
+ "type": "interline_equation",
+ "content": "\\mathrm {s} (Q, K _ {i}) = \\frac {e (Q) \\cdot e (K _ {i})}{\\| e (Q) \\| \\| e (K _ {i}) \\|}, \\tag {1}",
+ "image_path": "cc7b11eac7dd9c098a988bc8d1c6e280cb5f867cb0f49200d69284f82099eba3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 45,
+ 587,
+ 279,
+ 646
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 587,
+ 279,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 587,
+ 279,
+ 646
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 45,
+ 587,
+ 279,
+ 646
+ ],
+ "type": "inline_equation",
+ "content": "e(\\cdot)"
+ },
+ {
+ "bbox": [
+ 45,
+ 587,
+ 279,
+ 646
+ ],
+ "type": "text",
+ "content": " is the embedding function. If "
+ },
+ {
+ "bbox": [
+ 45,
+ 587,
+ 279,
+ 646
+ ],
+ "type": "inline_equation",
+ "content": "K_{i}"
+ },
+ {
+ "bbox": [
+ 45,
+ 587,
+ 279,
+ 646
+ ],
+ "type": "text",
+ "content": " has the highest similarity score to the given query "
+ },
+ {
+ "bbox": [
+ 45,
+ 587,
+ 279,
+ 646
+ ],
+ "type": "inline_equation",
+ "content": "Q"
+ },
+ {
+ "bbox": [
+ 45,
+ 587,
+ 279,
+ 646
+ ],
+ "type": "text",
+ "content": ", its corresponding step-by-step instruction "
+ },
+ {
+ "bbox": [
+ 45,
+ 587,
+ 279,
+ 646
+ ],
+ "type": "inline_equation",
+ "content": "I_{step}"
+ },
+ {
+ "bbox": [
+ 45,
+ 587,
+ 279,
+ 646
+ ],
+ "type": "text",
+ "content": " will be selected as the demonstration example "
+ },
+ {
+ "bbox": [
+ 45,
+ 587,
+ 279,
+ 646
+ ],
+ "type": "inline_equation",
+ "content": "D_{step}"
+ },
+ {
+ "bbox": [
+ 45,
+ 587,
+ 279,
+ 646
+ ],
+ "type": "text",
+ "content": " for the given high-level instruction "
+ },
+ {
+ "bbox": [
+ 45,
+ 587,
+ 279,
+ 646
+ ],
+ "type": "inline_equation",
+ "content": "I_{high}"
+ },
+ {
+ "bbox": [
+ 45,
+ 587,
+ 279,
+ 646
+ ],
+ "type": "text",
+ "content": "."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 45,
+ 653,
+ 255,
+ 665
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 653,
+ 255,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 653,
+ 255,
+ 665
+ ],
+ "type": "text",
+ "content": "4.2. Room-and-Object Aware Scene Perceiver"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 45,
+ 671,
+ 279,
+ 729
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 671,
+ 279,
+ 729
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 671,
+ 279,
+ 729
+ ],
+ "type": "text",
+ "content": "Though the world knowledge acquired from the static LLMs planner could benefit the embodied task promisingly, the static LLMs planner may generate wrong or irrelevant plannings, which misleads the agent. To address this issue, the LLM planner should be aware of and interact with the"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 296,
+ 106,
+ 530,
+ 303
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 296,
+ 106,
+ 530,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 296,
+ 106,
+ 530,
+ 303
+ ],
+ "type": "text",
+ "content": "dynamic observations. In [35], the names of objects obtained from the ground truth or pre-trained detectors have been added to in-context prompts. However, the agents of these works act in a very limited space with simple scenes and monotonous objects. By contrast, REVERIE involves large-area exploration between different floors and rooms, where the scenes are more complex with more diverse objects. Considering these factors, we propose a room-and-object aware scene perceiver (ROASP) for the LLM planner, which predicts not only the room type but also the visible objects of the current location. Rather than using separate classifiers and detectors to individually predict each position's room types and visible object categories, we use CLIP [30] as the proposed room-and-object aware scene perceiver. Thanks to CLIP's strong ability of zero-shot image classification in the open world, the ROASP can well handle these two tasks."
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 296,
+ 303,
+ 530,
+ 408
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 296,
+ 303,
+ 530,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 296,
+ 303,
+ 530,
+ 408
+ ],
+ "type": "text",
+ "content": "Specifically, we first fetch the room type labels from the MatterPort3D [6] semantic annotations and the object type labels are extracted from the REVERIE training dataset. They are used to build the codebook for the room categories "
+ },
+ {
+ "bbox": [
+ 296,
+ 303,
+ 530,
+ 408
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{C}_{\\mathrm{room}}"
+ },
+ {
+ "bbox": [
+ 296,
+ 303,
+ 530,
+ 408
+ ],
+ "type": "text",
+ "content": " and the object categories "
+ },
+ {
+ "bbox": [
+ 296,
+ 303,
+ 530,
+ 408
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{C}_{\\mathrm{obj}}"
+ },
+ {
+ "bbox": [
+ 296,
+ 303,
+ 530,
+ 408
+ ],
+ "type": "text",
+ "content": ", respectively. Then, at each timestep "
+ },
+ {
+ "bbox": [
+ 296,
+ 303,
+ 530,
+ 408
+ ],
+ "type": "inline_equation",
+ "content": "t"
+ },
+ {
+ "bbox": [
+ 296,
+ 303,
+ 530,
+ 408
+ ],
+ "type": "text",
+ "content": ", the agent perceives the environment and obtains the panoramic visual observation "
+ },
+ {
+ "bbox": [
+ 296,
+ 303,
+ 530,
+ 408
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{R}_t = \\{r_i\\}_{i=1}^n"
+ },
+ {
+ "bbox": [
+ 296,
+ 303,
+ 530,
+ 408
+ ],
+ "type": "text",
+ "content": ". For each single-view observation "
+ },
+ {
+ "bbox": [
+ 296,
+ 303,
+ 530,
+ 408
+ ],
+ "type": "inline_equation",
+ "content": "r_i"
+ },
+ {
+ "bbox": [
+ 296,
+ 303,
+ 530,
+ 408
+ ],
+ "type": "text",
+ "content": " in the panorama, the image feature "
+ },
+ {
+ "bbox": [
+ 296,
+ 303,
+ 530,
+ 408
+ ],
+ "type": "inline_equation",
+ "content": "f_{\\mathrm{r}}"
+ },
+ {
+ "bbox": [
+ 296,
+ 303,
+ 530,
+ 408
+ ],
+ "type": "text",
+ "content": " is extracted by the CLIP Image Encoder"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 381,
+ 417,
+ 530,
+ 431
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 381,
+ 417,
+ 530,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 381,
+ 417,
+ 530,
+ 431
+ ],
+ "type": "interline_equation",
+ "content": "f _ {\\mathrm {r}} = E _ {\\mathrm {C L I P}} ^ {v} \\left(r _ {i}\\right), \\tag {2}",
+ "image_path": "ba0bbe56e8b9e6e5ead645b9a1d42555a0ef4a68074c1e7ed16234e38ef19695.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 296,
+ 439,
+ 530,
+ 508
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 296,
+ 439,
+ 530,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 296,
+ 439,
+ 530,
+ 508
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 296,
+ 439,
+ 530,
+ 508
+ ],
+ "type": "inline_equation",
+ "content": "E_{\\mathrm{CLIP}}^v (\\cdot)"
+ },
+ {
+ "bbox": [
+ 296,
+ 439,
+ 530,
+ 508
+ ],
+ "type": "text",
+ "content": " represents the CLIP Image Encoder. For each room category "
+ },
+ {
+ "bbox": [
+ 296,
+ 439,
+ 530,
+ 508
+ ],
+ "type": "inline_equation",
+ "content": "c_{room}"
+ },
+ {
+ "bbox": [
+ 296,
+ 439,
+ 530,
+ 508
+ ],
+ "type": "text",
+ "content": " and each object category "
+ },
+ {
+ "bbox": [
+ 296,
+ 439,
+ 530,
+ 508
+ ],
+ "type": "inline_equation",
+ "content": "c_{obj}"
+ },
+ {
+ "bbox": [
+ 296,
+ 439,
+ 530,
+ 508
+ ],
+ "type": "text",
+ "content": ", we respectively construct a text phrase of room "
+ },
+ {
+ "bbox": [
+ 296,
+ 439,
+ 530,
+ 508
+ ],
+ "type": "inline_equation",
+ "content": "T_{\\mathrm{room}}"
+ },
+ {
+ "bbox": [
+ 296,
+ 439,
+ 530,
+ 508
+ ],
+ "type": "text",
+ "content": " as \"a photo of a "
+ },
+ {
+ "bbox": [
+ 296,
+ 439,
+ 530,
+ 508
+ ],
+ "type": "inline_equation",
+ "content": "\\{c_{room}\\}"
+ },
+ {
+ "bbox": [
+ 296,
+ 439,
+ 530,
+ 508
+ ],
+ "type": "text",
+ "content": "\" and a text phrase of object "
+ },
+ {
+ "bbox": [
+ 296,
+ 439,
+ 530,
+ 508
+ ],
+ "type": "inline_equation",
+ "content": "T_{\\mathrm{obj}}"
+ },
+ {
+ "bbox": [
+ 296,
+ 439,
+ 530,
+ 508
+ ],
+ "type": "text",
+ "content": " as \"a photo of a "
+ },
+ {
+ "bbox": [
+ 296,
+ 439,
+ 530,
+ 508
+ ],
+ "type": "inline_equation",
+ "content": "\\{c_{obj}\\}"
+ },
+ {
+ "bbox": [
+ 296,
+ 439,
+ 530,
+ 508
+ ],
+ "type": "text",
+ "content": "\". Then the text feature is derived through the pretrained CLIP Text Encoder as:"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 369,
+ 517,
+ 530,
+ 531
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 369,
+ 517,
+ 530,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 369,
+ 517,
+ 530,
+ 531
+ ],
+ "type": "interline_equation",
+ "content": "f _ {\\text {r o o m}} = E _ {\\text {C L I P}} ^ {t} \\left(T _ {\\text {r o o m}}\\right), \\tag {3}",
+ "image_path": "0a8fe8a3fe0ed28c72a7e3e66fa658cfa8f0a27107343cb099bf54d114a02e29.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 375,
+ 533,
+ 530,
+ 547
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 375,
+ 533,
+ 530,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 375,
+ 533,
+ 530,
+ 547
+ ],
+ "type": "interline_equation",
+ "content": "f _ {\\mathrm {o b j}} = E _ {\\mathrm {C L I P}} ^ {t} \\left(T _ {\\mathrm {o b j}}\\right), \\tag {4}",
+ "image_path": "4fcee2fbf2832236f8970a4ee23e720d5d740be004bd76f8bd677844d373ae02.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 296,
+ 554,
+ 530,
+ 613
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 296,
+ 554,
+ 530,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 296,
+ 554,
+ 530,
+ 613
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 296,
+ 554,
+ 530,
+ 613
+ ],
+ "type": "inline_equation",
+ "content": "E_{\\mathrm{CLIP}}^{t}(\\cdot)"
+ },
+ {
+ "bbox": [
+ 296,
+ 554,
+ 530,
+ 613
+ ],
+ "type": "text",
+ "content": " represents the CLIP Text Encoder. At last, the similarity score "
+ },
+ {
+ "bbox": [
+ 296,
+ 554,
+ 530,
+ 613
+ ],
+ "type": "inline_equation",
+ "content": "S_{\\mathrm{room}}"
+ },
+ {
+ "bbox": [
+ 296,
+ 554,
+ 530,
+ 613
+ ],
+ "type": "text",
+ "content": " between the image feature "
+ },
+ {
+ "bbox": [
+ 296,
+ 554,
+ 530,
+ 613
+ ],
+ "type": "inline_equation",
+ "content": "f_{\\mathrm{r}}"
+ },
+ {
+ "bbox": [
+ 296,
+ 554,
+ 530,
+ 613
+ ],
+ "type": "text",
+ "content": " and the text feature "
+ },
+ {
+ "bbox": [
+ 296,
+ 554,
+ 530,
+ 613
+ ],
+ "type": "inline_equation",
+ "content": "f_{\\mathrm{room}}"
+ },
+ {
+ "bbox": [
+ 296,
+ 554,
+ 530,
+ 613
+ ],
+ "type": "text",
+ "content": " as well as the similarity score "
+ },
+ {
+ "bbox": [
+ 296,
+ 554,
+ 530,
+ 613
+ ],
+ "type": "inline_equation",
+ "content": "S_{\\mathrm{obj}}"
+ },
+ {
+ "bbox": [
+ 296,
+ 554,
+ 530,
+ 613
+ ],
+ "type": "text",
+ "content": " between the image feature "
+ },
+ {
+ "bbox": [
+ 296,
+ 554,
+ 530,
+ 613
+ ],
+ "type": "inline_equation",
+ "content": "f_{\\mathrm{r}}"
+ },
+ {
+ "bbox": [
+ 296,
+ 554,
+ 530,
+ 613
+ ],
+ "type": "text",
+ "content": " and the text feature "
+ },
+ {
+ "bbox": [
+ 296,
+ 554,
+ 530,
+ 613
+ ],
+ "type": "inline_equation",
+ "content": "f_{\\mathrm{obj}}"
+ },
+ {
+ "bbox": [
+ 296,
+ 554,
+ 530,
+ 613
+ ],
+ "type": "text",
+ "content": " are respectively computed as:"
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 354,
+ 621,
+ 530,
+ 635
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 354,
+ 621,
+ 530,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 354,
+ 621,
+ 530,
+ 635
+ ],
+ "type": "interline_equation",
+ "content": "S _ {\\text {r o o m}} = \\operatorname {S o f t m a x} \\left(f _ {\\text {r o o m}} \\cdot f _ {\\mathrm {r}} ^ {T}\\right), \\tag {5}",
+ "image_path": "f1a20e2d3865d87ead58a3e4aaaaa3125b0ef369c8fa6911abcf64c91485f320.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 361,
+ 637,
+ 530,
+ 651
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 361,
+ 637,
+ 530,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 361,
+ 637,
+ 530,
+ 651
+ ],
+ "type": "interline_equation",
+ "content": "S _ {\\mathrm {o b j}} = \\operatorname {S o f t m a x} \\left(f _ {\\mathrm {o b j}} \\cdot f _ {\\mathrm {r}} ^ {T}\\right). \\tag {6}",
+ "image_path": "759fbb1b4a293fc0ead607c02f636969d37aaf79dcd68ad6f52cbf943bbb4619.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 296,
+ 660,
+ 530,
+ 730
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 296,
+ 660,
+ 530,
+ 730
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 296,
+ 660,
+ 530,
+ 730
+ ],
+ "type": "text",
+ "content": "Considering that the current environment normally belongs to only one type of room, though the panoramic images have multiple views, the room that the agent is currently centered in should have the largest influence on each view. Thus, we average the predicted room type scores "
+ },
+ {
+ "bbox": [
+ 296,
+ 660,
+ 530,
+ 730
+ ],
+ "type": "inline_equation",
+ "content": "S_{\\mathrm{room}}"
+ },
+ {
+ "bbox": [
+ 296,
+ 660,
+ 530,
+ 730
+ ],
+ "type": "text",
+ "content": " from multiple views and choose the room type with"
+ }
+ ]
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 277,
+ 746,
+ 300,
+ 756
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 277,
+ 746,
+ 300,
+ 756
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 277,
+ 746,
+ 300,
+ 756
+ ],
+ "type": "text",
+ "content": "15762"
+ }
+ ]
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "page_size": [
+ 595,
+ 842
+ ],
+ "page_idx": 4
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 52,
+ 110,
+ 272,
+ 184
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 52,
+ 110,
+ 272,
+ 184
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 52,
+ 110,
+ 272,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 52,
+ 110,
+ 272,
+ 184
+ ],
+ "type": "table",
+ "html": "| REVERIE Instruction: Go to the bathroom with the two towels and empty the lower cabinet. |
| Scene-Oriented Dynamic Planning | w/o Dynamic Demonstration | w/o ROASP |
| Step 1: exit the bedroom | Step 1: put the towel in the bottom drawer of the toilet | Step 1: exit the bathroom |
| Step 2: and walk into the next bedroom in the hall | Step 2: walk into the kitchen |
| Step 3: go stand in the kitchen doorway |
| Step 3: go stand in the bathroom | Step 4: go back to the bedroom and take the towel |
| Step 5: exit the bedroom |
| Step 6: walk into the next bedroom in the hall |
| REVERIE Instruction: Go to the kitchen on level 2 and open the drawer. |
| Scene-Oriented Dynamic Planning | w/o Dynamic Demonstration | w/o ROASP |
| Step 1: go upstairs | Step 1: open the drawer with the key in it | Step 1: go upstairs and turn left at the hall |
| Step 2: go straight pass the couch | Step 2: remove the lock from the door | Step 2: go back down the stairs |
| Step 3: make a slight left towards the refrigerator | Step 3: then turn on | Step 3: go up the stairs again |
| Step 4: wait near the kitchen drawer | Step 4: go back to the kitchen and open the fridge |
| Step 5: take out the fruit and put it in the juicer |
",
+ "image_path": "388916c8cba04ced0696f35036aed1758ff7903a6c11a0c3aaa5c5c89f16678c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 198,
+ 303,
+ 377,
+ 314
+ ],
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 198,
+ 303,
+ 377,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 198,
+ 303,
+ 377,
+ 314
+ ],
+ "type": "text",
+ "content": "Figure 6: Examples of generated instructions."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 45,
+ 321,
+ 281,
+ 484
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 321,
+ 281,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 321,
+ 281,
+ 484
+ ],
+ "type": "text",
+ "content": "fine-grained static instruction and scene-oriented dynamic instruction is that static fine-grained instruction is generated without ROASP. More specifically, the query prompt for the LLM to generate step-by-step planning is fixed at each timestep, which only consists of the given high-level instruction and the selected demonstration. (III) Dynamic: The input assembled instruction contains the high-level instruction in REVERIE and scene-oriented dynamic planning instruction. As shown in Table 3, in the static setting, the performance in all metrics is improved compared to the baseline, indicating the effectiveness of the LLM's rich world knowledge in fine-grained planning. In the dynamic setting, the performance is further improved with non-trivial margins, showing the effectiveness of ROASP."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 45,
+ 497,
+ 279,
+ 706
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 497,
+ 279,
+ 706
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 497,
+ 279,
+ 706
+ ],
+ "type": "text",
+ "content": "Qualitative Analysis of Prompt Setting To further evaluate the effect of dynamic demonstration and ROASP in SODP, we perform a human evaluation about the generated plannings (see Table 4) and show the planning results (see Fig. 6). For human evaluation, we randomly selected 100 REVERIE tasks and generate fine-grained step-by-step instructions in setting of SODP, SODP without dynamic demonstration, and SODP without ROASP. We asked 10 volunteers to mark the generated step-by-step instructions in terms of their relevancy and rationality. The relevancy score ranges from 0 (unrelated) to 3 (very related), which takes into account whether the keywords in instructions are related to the REVERIE task. For example, regarding the REVERIE instruction \"Go to the kitchen and turn on the microwave\", whether there are keywords in instructions related to the kitchen scene could be rated. Rationality is rated from 0 (bad) to 3 (perfect), considering whether the instruction conforms to the logic of navigation."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 45,
+ 706,
+ 279,
+ 730
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 706,
+ 279,
+ 730
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 706,
+ 279,
+ 730
+ ],
+ "type": "text",
+ "content": "The results are presented in Table 4. It shows that our SODP scored 2.06 on Relevancy and 1.93 on Rational-"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 296,
+ 321,
+ 530,
+ 519
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 296,
+ 321,
+ 530,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 296,
+ 321,
+ 530,
+ 519
+ ],
+ "type": "text",
+ "content": "ity, which could be considered acceptable since the highest score is 3 and it is challenging to generate instructions that are consistent with tasks and actual navigation logic. When removing the dynamic demonstration, the score of generated instruction drops about "
+ },
+ {
+ "bbox": [
+ 296,
+ 321,
+ 530,
+ 519
+ ],
+ "type": "inline_equation",
+ "content": "31.55\\%"
+ },
+ {
+ "bbox": [
+ 296,
+ 321,
+ 530,
+ 519
+ ],
+ "type": "text",
+ "content": " on Relevancy and "
+ },
+ {
+ "bbox": [
+ 296,
+ 321,
+ 530,
+ 519
+ ],
+ "type": "inline_equation",
+ "content": "36.27\\%"
+ },
+ {
+ "bbox": [
+ 296,
+ 321,
+ 530,
+ 519
+ ],
+ "type": "text",
+ "content": " on Rationality, which could also be observed in Fig. 6. Although the instruction generated without dynamic demonstration is related to the task to some extent (e.g., \"put the towel in the bottom drawer of the toilet\" has the keyword \"towel\", the instruction lacks navigation information, such as how to reach the bathroom.) As shown in the bottom example of Fig. 6, instruction without ROASP successfully guided how to go to the destination location kitchen, but it still caused confusion by going upstairs and going downstairs several times, and thus reducing the rationality score, i.e. 1.64 on Relevancy and 1.55 on Rationality. More generation results can be found in the supplementary."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 297,
+ 528,
+ 373,
+ 541
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 297,
+ 528,
+ 373,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 297,
+ 528,
+ 373,
+ 541
+ ],
+ "type": "text",
+ "content": "6. Conclusions"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 296,
+ 548,
+ 530,
+ 676
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 296,
+ 548,
+ 530,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 296,
+ 548,
+ 530,
+ 676
+ ],
+ "type": "text",
+ "content": "In this work, we propose a novel model, March-in-Chat (MiC), for the REVEIRE task, which only provides concise high-level instructions for the VLN agent. MiC enables the REVERIE agent to talk with an LLM on the fly to generate plans for the next few steps. It consists of three main modules, Goal-Oriented Static Planning (GOSP), Scene-Oriented Dynamic Planning (SODP), and Room-and-Object Aware Scene Perceiver (ROASP) module. We conduct extensive quantitative and qualitative experiments on REVERIE and the promising results show the effectiveness of our method."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 297,
+ 687,
+ 402,
+ 700
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 297,
+ 687,
+ 402,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 297,
+ 687,
+ 402,
+ 700
+ ],
+ "type": "text",
+ "content": "7. Acknowledgements"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 297,
+ 706,
+ 530,
+ 730
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 297,
+ 706,
+ 530,
+ 730
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 297,
+ 706,
+ 530,
+ 730
+ ],
+ "type": "text",
+ "content": "Jing Liu is supported by the National Key Research and Development Program of China (No. 2020AAA0106400)."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 277,
+ 746,
+ 300,
+ 756
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 277,
+ 746,
+ 300,
+ 756
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 277,
+ 746,
+ 300,
+ 756
+ ],
+ "type": "text",
+ "content": "15765"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "page_size": [
+ 595,
+ 842
+ ],
+ "page_idx": 7
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 47,
+ 105,
+ 103,
+ 117
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 105,
+ 103,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 105,
+ 103,
+ 117
+ ],
+ "type": "text",
+ "content": "References"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 51,
+ 124,
+ 279,
+ 729
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 10,
+ "blocks": [
+ {
+ "bbox": [
+ 51,
+ 124,
+ 279,
+ 273
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 124,
+ 279,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 124,
+ 279,
+ 273
+ ],
+ "type": "text",
+ "content": "[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Daniel Ho, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Eric Jang, Rosario Jauregui Ruano, Kyle Jeffrey, Sally Jesmonth, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Kuang-Huei Lee, Sergey Levine, Yao Lu, Linda Luu, Carolina Parada, Peter Pastor, Jornell Quiambao, Kanishka Rao, Jarek Rettinghouse, Diego Reyes, Pierre Sermanet, Nicolas Sievers, Clayton Tan, Alexander Toshev, Vincent Vanhoucke, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Mengyuan Yan, and Andy Zeng. Do as i can and not as i say: Grounding language in robotic affordances. In arXiv preprint arXiv:2204.01691, 2022. 3"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 51,
+ 275,
+ 279,
+ 317
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 275,
+ 279,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 275,
+ 279,
+ 317
+ ],
+ "type": "text",
+ "content": "[2] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Topo-metric map pre-training for language-guided navigation. arXiv preprint arXiv:2212.04385, 2022. 2"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 51,
+ 319,
+ 279,
+ 372
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 319,
+ 279,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 319,
+ 279,
+ 372
+ ],
+ "type": "text",
+ "content": "[3] Peter Anderson, Angel X. Chang, Devendra Singh Chaplot, Alexey Dosovitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, and Amir Roshan Zamir. On evaluation of embodied navigation agents. CoRR, abs/1807.06757, 2018. 6"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 51,
+ 373,
+ 279,
+ 426
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 373,
+ 279,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 373,
+ 279,
+ 426
+ ],
+ "type": "text",
+ "content": "[4] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, pages 3674-3683, 2018. 1, 2, 5"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 51,
+ 427,
+ 279,
+ 533
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 427,
+ 279,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 427,
+ 279,
+ 533
+ ],
+ "type": "text",
+ "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 3, 4"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 51,
+ 535,
+ 279,
+ 587
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 535,
+ 279,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 535,
+ 279,
+ 587
+ ],
+ "type": "text",
+ "content": "[6] Angel X. Chang, Angela Dai, Thomas A. Funkhouser, Macej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from RGB-D data in indoor environments. In 3DV, pages 667-676, 2017. 5"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 51,
+ 589,
+ 279,
+ 631
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 589,
+ 279,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 589,
+ 279,
+ 631
+ ],
+ "type": "text",
+ "content": "[7] Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. TOUCHDOWN: natural language navigation and spatial reasoning in visual street environments. In CVPR, pages 12538-12547, 2019. 1, 2"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 51,
+ 633,
+ 279,
+ 685
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 633,
+ 279,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 633,
+ 279,
+ 685
+ ],
+ "type": "text",
+ "content": "[8] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021. 3"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 51,
+ 687,
+ 279,
+ 729
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 687,
+ 279,
+ 729
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 687,
+ 279,
+ 729
+ ],
+ "type": "text",
+ "content": "[9] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, pages 5834-5847, 2021. 7"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "sub_type": "ref_text"
+ },
+ {
+ "bbox": [
+ 299,
+ 107,
+ 530,
+ 729
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 24,
+ "blocks": [
+ {
+ "bbox": [
+ 299,
+ 107,
+ 530,
+ 149
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 107,
+ 530,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 107,
+ 530,
+ 149
+ ],
+ "type": "text",
+ "content": "[10] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Learning from unlabeled 3d environments for vision-and-language navigation. In ECCV, 2022. 4, 6, 7"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 299,
+ 151,
+ 530,
+ 193
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 151,
+ 530,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 151,
+ 530,
+ 193
+ ],
+ "type": "text",
+ "content": "[11] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, 2022. 2, 7"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 299,
+ 195,
+ 530,
+ 237
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 195,
+ 530,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 195,
+ 530,
+ 237
+ ],
+ "type": "text",
+ "content": "[12] Chen Gao, Jinyu Chen, Si Liu, Luting Wang, Qiong Zhang, and Qi Wu. Room-and-object aware knowledge reasoning for remote embodied referring expression. In CVPR, pages 3064-3073, 2021. 7"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 299,
+ 239,
+ 530,
+ 280
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 239,
+ 530,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 239,
+ 530,
+ 280
+ ],
+ "type": "text",
+ "content": "[13] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In ICCV, pages 1634-1643, 2021. 7"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 299,
+ 282,
+ 530,
+ 325
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 282,
+ 530,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 282,
+ 530,
+ 325
+ ],
+ "type": "text",
+ "content": "[14] Keji He, Yan Huang, Qi Wu, Jianhua Yang, Dong An, Shuanglin Sima, and Liang Wang. Landmark-rxr: Solving vision-and-language navigation with fine-grained alignment supervision. NeurIPS, 34:652-663, 2021. 2"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 299,
+ 326,
+ 530,
+ 369
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 326,
+ 530,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 326,
+ 530,
+ 369
+ ],
+ "type": "text",
+ "content": "[15] Yicong Hong, Cristian Rodriguez Opazo, Qi Wu, and Stephen Gould. Sub-instruction aware vision-and-language navigation. In Bonnie Webber, Trevor Cohn, Yulan He, and Yang Liu, editors, EMNLP, pages 3360-3376, 2020. 5"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 299,
+ 370,
+ 530,
+ 412
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 370,
+ 530,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 370,
+ 530,
+ 412
+ ],
+ "type": "text",
+ "content": "[16] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez Opazo, and Stephen Gould. VLN"
+ },
+ {
+ "bbox": [
+ 299,
+ 370,
+ 530,
+ 412
+ ],
+ "type": "inline_equation",
+ "content": "\\text{©}"
+ },
+ {
+ "bbox": [
+ 299,
+ 370,
+ 530,
+ 412
+ ],
+ "type": "text",
+ "content": " BERT: A recurrent vision-and-language BERT for navigation. In CVPR, pages 1643-1653, 2021. 2, 7"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 299,
+ 414,
+ 530,
+ 457
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 414,
+ 530,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 414,
+ 530,
+ 457
+ ],
+ "type": "text",
+ "content": "[17] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In ICML, pages 9118-9147. PMLR, 2022. 2, 3, 5"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 299,
+ 458,
+ 530,
+ 532
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 458,
+ 530,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 458,
+ 530,
+ 532
+ ],
+ "type": "text",
+ "content": "[18] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, Pierre Sermanet, Noah Brown, Tomas Jackson, Linda Luu, Sergey Levine, Karol Hausman, and Brian Ichter. Inner monologue: Embodied reasoning through planning with language models. In arXiv preprint arXiv:2207.05608, 2022. 3"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 299,
+ 533,
+ 530,
+ 576
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 533,
+ 530,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 533,
+ 530,
+ 576
+ ],
+ "type": "text",
+ "content": "[19] Vihan Jain, Gabriel Magalhães, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. pages 1862-1872, 2019. 2"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 299,
+ 578,
+ 530,
+ 641
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 578,
+ 530,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 578,
+ 530,
+ 641
+ ],
+ "type": "text",
+ "content": "[20] Melvin Johnson, Mike Schuster, Quoc V. Le, Maxim Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Thorat, Fernanda B. Viégas, Martin Wattenberg, Greg Corrado, Macduff Hughes, and Jeffrey Dean. Google's multilingual neural machine translation system: Enabling zero-shot translation. Trans. Assoc. Comput. Linguistics, 5:339-351, 2017. 6"
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 299,
+ 643,
+ 530,
+ 685
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 643,
+ 530,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 643,
+ 530,
+ 685
+ ],
+ "type": "text",
+ "content": "[21] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In EMNLP, pages 4392-4412, 2020. 1, 2"
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 299,
+ 687,
+ 530,
+ 729
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 687,
+ 530,
+ 729
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 687,
+ 530,
+ 729
+ ],
+ "type": "text",
+ "content": "[22] Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. What makes good in-context examples for gpt-3? arXiv preprint arXiv:2101.06804, 2021. 5"
+ }
+ ]
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 277,
+ 747,
+ 300,
+ 756
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 277,
+ 747,
+ 300,
+ 756
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 277,
+ 747,
+ 300,
+ 756
+ ],
+ "type": "text",
+ "content": "15766"
+ }
+ ]
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "page_size": [
+ 595,
+ 842
+ ],
+ "page_idx": 8
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 47,
+ 106,
+ 279,
+ 730
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 13,
+ "blocks": [
+ {
+ "bbox": [
+ 47,
+ 106,
+ 279,
+ 149
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 106,
+ 279,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 106,
+ 279,
+ 149
+ ],
+ "type": "text",
+ "content": "[23] Chih-Yao Ma, Jiasen Lu, Zuxuan Wu, Ghassan AlRegib, Zsolt Kira, Richard Socher, and Caiming Xiong. Self-monitoring navigation agent via auxiliary progress estimation. In ICLR, 2019. 7"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 47,
+ 151,
+ 279,
+ 194
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 151,
+ 279,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 151,
+ 279,
+ 194
+ ],
+ "type": "text",
+ "content": "[24] Khanh Nguyen and Hal Daumé III. Help, anna! visual navigation with natural multimodal assistance via retrospective curiosity-encouraging imitation learning. pages 684–695. Association for Computational Linguistics, 2019. 2"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 47,
+ 195,
+ 279,
+ 237
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 195,
+ 279,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 195,
+ 279,
+ 237
+ ],
+ "type": "text",
+ "content": "[25] Gabriel Poesia, Alex Polozov, Vu Le, Ashish Tiwari, Gustavo Soares, Christopher Meek, and Sumit Gulwani. Synchromesh: Reliable code generation from pre-trained language models. In ICLR, 2022. 5"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 47,
+ 239,
+ 279,
+ 292
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 239,
+ 279,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 239,
+ 279,
+ 292
+ ],
+ "type": "text",
+ "content": "[26] Yuankai Qi, Zizheng Pan, Yicong Hong, Ming-Hsuan Yang, Anton van den Hengel, and Qi Wu. The road to knowwhere: An object-and-room informed sequential bert for indoor vision-language navigation. In ICCV, pages 1655-1664, 2021. 2, 7"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 47,
+ 293,
+ 279,
+ 346
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 293,
+ 279,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 293,
+ 279,
+ 346
+ ],
+ "type": "text",
+ "content": "[27] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. REVERIE: remote embodied visual referring expression in real indoor environments. In CVPR, pages 9979-9988, 2020. 1, 2, 6, 7"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 47,
+ 347,
+ 279,
+ 389
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 347,
+ 279,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 347,
+ 279,
+ 389
+ ],
+ "type": "text",
+ "content": "[28] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. HOP: history-and-order aware pretraining for vision-and-language navigation. In CVPR, pages 15397-15406, 2022. 2, 7"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 47,
+ 391,
+ 279,
+ 433
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 391,
+ 279,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 391,
+ 279,
+ 433
+ ],
+ "type": "text",
+ "content": "[29] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. Hop+: History-enhanced and order-aware pre-training for vision-and-language navigation. IEEE TPAMI, 2023. 2"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 47,
+ 435,
+ 279,
+ 500
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 435,
+ 279,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 435,
+ 279,
+ 500
+ ],
+ "type": "text",
+ "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, volume 139, pages 8748-8763, 2021. 5"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 47,
+ 500,
+ 279,
+ 542
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 500,
+ 279,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 500,
+ 279,
+ 542
+ ],
+ "type": "text",
+ "content": "[31] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3, 6"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 47,
+ 544,
+ 279,
+ 586
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 544,
+ 279,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 544,
+ 279,
+ 586
+ ],
+ "type": "text",
+ "content": "[32] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors, EMNLP-IJCNLP, pages 3980-3990, 2019. 5"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 47,
+ 588,
+ 279,
+ 640
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 588,
+ 279,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 588,
+ 279,
+ 640
+ ],
+ "type": "text",
+ "content": "[33] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A benchmark for interpreting grounded instructions for everyday tasks. In CVPR, pages 10737-10746, 2020. 1, 3"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 47,
+ 642,
+ 279,
+ 685
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 642,
+ 279,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 642,
+ 279,
+ 685
+ ],
+ "type": "text",
+ "content": "[34] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. ProgPrompt: Generating situated robot task plans using large language models. 2022. 2"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 47,
+ 687,
+ 279,
+ 730
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 687,
+ 279,
+ 730
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 687,
+ 279,
+ 730
+ ],
+ "type": "text",
+ "content": "[35] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. arXiv preprint arXiv:2212.04088, 2022. 3, 5"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "sub_type": "ref_text"
+ },
+ {
+ "bbox": [
+ 299,
+ 106,
+ 530,
+ 280
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 18,
+ "blocks": [
+ {
+ "bbox": [
+ 299,
+ 106,
+ 530,
+ 138
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 106,
+ 530,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 106,
+ 530,
+ 138
+ ],
+ "type": "text",
+ "content": "[36] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In CoRL, pages 394-406, 2019. 1, 2"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 299,
+ 140,
+ 530,
+ 193
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 140,
+ 530,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 140,
+ 530,
+ 193
+ ],
+ "type": "text",
+ "content": "[37] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In CVPR, pages 6629-6638, 2019. 7"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 299,
+ 194,
+ 530,
+ 236
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 194,
+ 530,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 194,
+ 530,
+ 236
+ ],
+ "type": "text",
+ "content": "[38] Yusheng Zhao, Jinyu Chen, Chen Gao, Wenguan Wang, Lirong Yang, Haibing Ren, Huaxia Xia, and Si Liu. Target-driven structured transformer planner for vision-language navigation. pages 4194-4203, 2022. 7"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 299,
+ 237,
+ 530,
+ 280
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 237,
+ 530,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 237,
+ 530,
+ 280
+ ],
+ "type": "text",
+ "content": "[39] Fengda Zhu, Xiwen Liang, Yi Zhu, Qizhi Yu, Xiaojun Chang, and Xiaodan Liang. SOON: scenario oriented object navigation with graph-based exploration. In CVPR, pages 12689-12699, 2021. 1, 2"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 277,
+ 746,
+ 300,
+ 756
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 277,
+ 746,
+ 300,
+ 756
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 277,
+ 746,
+ 300,
+ 756
+ ],
+ "type": "text",
+ "content": "15767"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "page_size": [
+ 595,
+ 842
+ ],
+ "page_idx": 9
+ }
+ ],
+ "_backend": "vlm",
+ "_version_name": "2.6.4"
+}
\ No newline at end of file
diff --git a/2023/Markov Game Video Augmentation for Action Segmentation/3ebb7b09-a6e7-4446-89ff-eb9741fd0b7f_content_list.json b/2023/Markov Game Video Augmentation for Action Segmentation/3ebb7b09-a6e7-4446-89ff-eb9741fd0b7f_content_list.json
new file mode 100644
index 0000000000000000000000000000000000000000..0137158ca2b098464b3d0bce8625bc2e79f37fb0
--- /dev/null
+++ b/2023/Markov Game Video Augmentation for Action Segmentation/3ebb7b09-a6e7-4446-89ff-eb9741fd0b7f_content_list.json
@@ -0,0 +1,1498 @@
+[
+ {
+ "type": "text",
+ "text": "Markov Game Video Augmentation for Action Segmentation",
+ "text_level": 1,
+ "bbox": [
+ 179,
+ 130,
+ 790,
+ 152
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Nicolas Aziere",
+ "bbox": [
+ 292,
+ 181,
+ 415,
+ 196
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Oregon State University",
+ "bbox": [
+ 256,
+ 199,
+ 450,
+ 217
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "azieren@oregonstate.edu",
+ "bbox": [
+ 251,
+ 219,
+ 457,
+ 232
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Sinisa Todorovic",
+ "bbox": [
+ 550,
+ 181,
+ 687,
+ 196
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Oregon State University",
+ "bbox": [
+ 522,
+ 199,
+ 714,
+ 217
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "sinisa@oregonstate.edu",
+ "bbox": [
+ 521,
+ 219,
+ 718,
+ 233
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Abstract",
+ "text_level": 1,
+ "bbox": [
+ 233,
+ 268,
+ 313,
+ 284
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "This paper addresses data augmentation for action segmentation. Our key novelty is that we augment the original training videos in the deep feature space, not in the visual spatiotemporal domain as done by previous work. For augmentation, we modify original deep features of video frames such that the resulting embeddings fall closer to the class decision boundaries. Also, we edit action sequences of the original training videos (a.k.a. transcripts) by inserting, deleting, and replacing actions such that the resulting transcripts are close in edit distance to the ground-truth ones. For our data augmentation we resort to reinforcement learning, instead of more common supervised learning, since we do not have access to reliable oracles which would provide supervision about the optimal data modifications in the deep feature space. For modifying frame embeddings, we use a meta-model formulated as a Markov Game with multiple self-interested agents. Also, new transcripts are generated using a fast, parameter-free Monte Carlo tree search. Our experiments show that the proposed data augmentation of the Breakfast, GTEA, and 50Salads datasets leads to significant performance gains of several state of the art action segmenters.",
+ "bbox": [
+ 75,
+ 301,
+ 472,
+ 632
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "1. Introduction",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 662,
+ 209,
+ 679
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "This paper presents a new data augmentation framework for fully supervised action segmentation of untrimmed videos. Action segmentation is a basic vision problem. Despite recent tremendous advances in terms of new action segmenters and learning strategies, there is relatively slow progress in increasing the size of existing benchmark datasets. In comparison with peer datasets for action recognition or image classification, available benchmarks for action segmentation are significantly smaller. This presents challenges in training of recent action segmenters which show tendency to overfit on small datasets [11, 43, 40]. However, compiling large datasets is difficult, due to, in part, high costs of manual annotation of action segments.",
+ "bbox": [
+ 75,
+ 688,
+ 468,
+ 885
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "We propose to augment existing datasets with newly gen",
+ "bbox": [
+ 96,
+ 886,
+ 468,
+ 901
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "erated video sequences, such that the resulting data augmentation enables more robust training and hence improves performance of action segmenters. Our approach is agnostic of a particular model for action segmentation, and expects that the segmenter has been pre-trained on the original training dataset to predict action classes of video frames.",
+ "bbox": [
+ 496,
+ 268,
+ 890,
+ 361
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Augmentation of video data has been mostly considered in the spatiotemporal, visual domain [38, 25, 19], where a human expert would heuristically specify the amount and type of data manipulation (e.g., subsampling, cropping, flipping of video frames) that are useful in training. While these approaches show great success, it is hard to formalize them in a principled manner. Others learn to generate new videos [9, 8, 42, 41], but the results are not sufficiently realistic yet, and hence would require domain adaptation if used for data augmentation in action segmentation.",
+ "bbox": [
+ 496,
+ 364,
+ 892,
+ 516
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Our key novelty is that we augment the original training videos directly in the deep feature space, unlike most previous work. As shown in Fig. 1 (top left), for augmentation, we modify original deep features of video frames at the input, such that the resulting embeddings fall closer to the class decision boundaries. Thus, by construction, we enforce that the augmented features be more challenging for learning, and in this way subsequently enable more robust training of the action segmenter. In addition, Fig. 1 (top right) illustrates that we also edit action sequences of the original training videos (a.k.a. transcripts) by inserting, deleting, and replacing actions, such that the resulting transcripts are close in edit distance to the groundtruth ones. Since the generated transcripts are kept similar to the originals, they are expected to be meaningful (i.e., legal) and provide a greater variety of legal action sequences than seen in the original training set. This is especially important for those application domains where some transcripts of interest are naturally rare and hence underrepresented in the original training dataset.",
+ "bbox": [
+ 496,
+ 518,
+ 893,
+ 821
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "For the proposed augmentation of frame embeddings in the deep feature space, we specify a deep residual metamodel, as shown in Fig. 1 (bottom left). The meta-model takes deep features of frames at the input and predicts an optimal amount of feature modifications - i.e., offset fea",
+ "bbox": [
+ 496,
+ 824,
+ 893,
+ 902
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "header",
+ "text": "CVF",
+ "bbox": [
+ 106,
+ 2,
+ 181,
+ 42
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "header",
+ "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+ "bbox": [
+ 238,
+ 0,
+ 805,
+ 46
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "page_number",
+ "text": "13505",
+ "bbox": [
+ 480,
+ 944,
+ 517,
+ 957
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "image",
+ "img_path": "images/07e320007427c9b028287eb228d18bdca74202a43981638ece08e9a5d1256c40.jpg",
+ "image_caption": [
+ "Figure 1. The proposed two-pronged data augmentation: (top and bottom left) Original deep features of frames at the input of a meta-model are modified to fall closer to class decision boundaries of the pretrained action segmenter. Alternative frameworks are considered for training the meta-model, including Gaussian noise, or supervised learning of a Temporal CNN, or reinforcement learning of an Actor-Critic network. (top and bottom right) The original action transcripts of training videos are modified by inserting, replacing, and removing actions, while ensuring a small edit distance between the newly generated transcripts and the originals. For training the meta-model for transcript augmentation, we consider supervised learning and alternatively a Monte Carlo tree search."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 81,
+ 90,
+ 467,
+ 251
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "tures which are added to the original frame features, resulting in augmented embeddings. For training the meta-model, we consider both RL (Fig. 1 (bottom left)) and supervised learning. From our experiments, RL proves to be a more reliable paradigm.",
+ "bbox": [
+ 75,
+ 446,
+ 467,
+ 521
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Supervised learning of the proposed meta-model faces a fundamental challenge. As a direct consequence of performing data augmentation in the deep feature space, there is no reliable oracle which would provide supervision for the optimal amount of feature modifications. The only available oracle is the pre-trained action segmentation model which could predict action labels of the augmented frame embeddings, and the incurred loss could be used for training the meta-model. The action segmentation model provides only pseudo-labels for generated data – not ground-truth labels – and hence a fully-supervised training of our meta-model could suffer from noisy pseudo-labels.",
+ "bbox": [
+ 75,
+ 522,
+ 467,
+ 703
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "An alternative is to resort to reinforcement learning (RL) due to the following advantages. First, it allows us to sequentially modify frame features, where previous modifications define a state in which RL estimates an optimal modification of the next frames. Learning of the proposed sequential feature augmentation is expected to be more reliable, especially for long video sequences, than learning how to optimally modify all frame embeddings at once. RL produces an optimal policy which would account for temporal dependence between frame embeddings, and hence make them suitable for action segmentation models with large temporal receptive fields (like MS-TCN or AS-Former). Second, RL is expected to provide for a more sta",
+ "bbox": [
+ 75,
+ 704,
+ 467,
+ 900
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "ble training of our meta-model by optimizing the expected reward over a policy of feature modifications, in comparison with the aforementioned minimization of the unreliable loss of the pretrained action segmenter on particular feature modifications. Third, RL is known to be very effective for problems with a large, continuous, output space, as is our case of predicting offset features in the deep feature space. Finally, RL is known to successfully address non-stationary environments with the distribution shift between training and test sets [37], which exactly characterizes our problem statement where data augmentation is aimed at bridging the distribution shift.",
+ "bbox": [
+ 496,
+ 90,
+ 890,
+ 271
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Within the RL framework, for modifying frame embeddings, we formulate the meta-model as a Deep Actor-Critic Network for learning policies of two self-interested agents in a Markov Game. Also, for generating new transcripts, we use a fast, parameter-free Monte Carlo tree search. We call our approach Markov Game Video Augmentation (MVGA).",
+ "bbox": [
+ 496,
+ 272,
+ 890,
+ 377
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Our experimental evaluation shows significant performance gains of recent convolutional and transformer-based action segmenters when our MVGA is used to augment the Breakfast, GTEA, and 50Salads datasets. Interestingly, MVGA enables the convolutional model MS-TCN [11, 27] achieve close performance to that of the significantly more complex (and more recent) ASFormer [43].",
+ "bbox": [
+ 496,
+ 378,
+ 890,
+ 483
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "In the following, Sec. 2 reviews closely related work, Sec. 3 gives an overview of MVGA, Sec. 4 specifies our transcript augmentation, Sec. 5 formalizes our frame-feature augmentation, and Sec. 6 presents our results.",
+ "bbox": [
+ 496,
+ 483,
+ 890,
+ 544
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2. Related Work",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 558,
+ 640,
+ 573
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "This section reviews closely related work on fully supervised action segmentation and video data augmentation.",
+ "bbox": [
+ 496,
+ 583,
+ 888,
+ 613
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "A family of temporal convolution models [10, 23, 24, 11, 40, 27, 34] have been studied for action segmentation. Some of these models use gradual temporal pooling for overcoming oversegmentation [10, 23, 24, 34], and others integrate reasoning about action boundaries [40, 18]. As representatives of this family, in our experiments, we consider MS-TCN [11, 27] and boundary-aware BCN [40] both of which consist of multiple stages of temporal convolution layers. Recent transformer-based models [43, 31, 39, 2] outperform temporal convolution models. Among these transformers, for evaluation, we consider ASFormer [43]. Both MS-TCN and ASFormer use standard I3D deep features [5]) of video frames as input. We also study a version of ASFormer with more informative frame embeddings [26].",
+ "bbox": [
+ 496,
+ 613,
+ 890,
+ 824
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Video augmentation in the visible space-time domain, such as, e.g., temporal cropping of frames [38, 25] or random cropping/removing/flipping of frames in a training mini-batch [19] have become standard practice in action recognition, person re-ID, and hand gesture recogni",
+ "bbox": [
+ 496,
+ 825,
+ 890,
+ 900
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "page_number",
+ "text": "13506",
+ "bbox": [
+ 480,
+ 944,
+ 519,
+ 955
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "tion. Other data augmentation methods have also been used, including pooling frame features within a window of variable length [34]. For video augmentation, some approaches generate simulated videos with video-game engines [9, 8], or GANs [14, 42, 41, 33]. However, there is still a large domain gap between such simulated and real videos, which limits the utility of these methods for data augmentation.",
+ "bbox": [
+ 75,
+ 90,
+ 467,
+ 196
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Reinforcement learning has been used for addressing various vision problems, including 3D image segmentation [28, 32], image classification [29], object detection [3, 4], and tracking [6, 16]. To the best of our knowledge, video data augmentation for action segmentation has never been formulated within the reinforcement learning framework.",
+ "bbox": [
+ 75,
+ 196,
+ 467,
+ 287
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "3. Overview of MVGA",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 303,
+ 267,
+ 318
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "In this and following two sections, we focus on our reinforcement learning formulation of the proposed data augmentation. The supervised learning formulation is described in Sec. 6. Fig. 2 shows an overview of our approach which consists of the following four steps. The first step pre-trains an action segmenter on a given training dataset $\\mathcal{D}$ . A training video of length $T$ in $\\mathcal{D}$ is given by its deep features for every frame $\\mathcal{X} = \\{x_t : t = 1, \\dots, T\\}$ , where $x_t \\in \\mathbb{R}^{d_{\\mathrm{in}}}$ ( $d_{\\mathrm{in}} = 2048$ for standard I3D features [5]), and the ground-truth action classes $\\mathcal{Y} = \\{y_t : t = 1, \\dots, T\\}$ , where $y_t \\in \\mathbb{Y}$ and $\\mathbb{Y}$ is a set of action classes. After pretraining, the action segmenter $f_{\\theta}$ can be used to predict action classes of video frames $\\hat{\\mathcal{Y}} = f(\\mathcal{X}; \\theta)$ .",
+ "bbox": [
+ 75,
+ 329,
+ 467,
+ 525
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "The second step generates new transcripts and their corresponding videos. We use the UCT algorithm [20] to first efficiently construct a tree, whose paths from the root to leaves represent legal transcripts, and then select optimal paths. For every new transcript, a new video is constructed by copying appropriate action segments from real videos in $\\mathcal{D}$ to the new video following the action sequence of the new transcript.",
+ "bbox": [
+ 75,
+ 526,
+ 467,
+ 645
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "The third step augments features of the original and new videos with the Actor and Critic networks, which gives the augmented dataset $\\mathcal{D}'$ . The Actor predicts how much and where to modify features in the input video. The Critic estimates the expected rewards for the Actor. Both Actor and Critic are learned using the pre-trained $f_{\\theta}$ as oracle.",
+ "bbox": [
+ 75,
+ 647,
+ 467,
+ 737
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "The fourth step fine-tunes $f_{\\theta}$ such that every mini-batch consists of videos from both $\\mathcal{D}$ and $\\mathcal{D}'$ .",
+ "bbox": [
+ 75,
+ 738,
+ 467,
+ 767
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "In the following, we specify the second and third steps.",
+ "bbox": [
+ 96,
+ 768,
+ 460,
+ 784
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "4. New Transcript and Video Generation",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 799,
+ 421,
+ 815
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "From $\\mathcal{D}$ , we generate new transcripts and their corresponding new videos. The new transcripts should be semantically meaningful, i.e., legal. This is enforced by requiring that the new transcripts have: (i) similar lengths as the original transcripts; and (ii) high likelihoods of consec",
+ "bbox": [
+ 75,
+ 825,
+ 467,
+ 900
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "utive pairs of action classes. Our experiments suggest that auto-regressive models - e.g., a recurrent-neural network (RNN) or Transformer network [35] - provide poor transcript generation, since they require large training datasets and hence have limited utility for our target settings where data augmentation is needed to address lack of data. Therefore, in this section, we focus on an alternative framework - parameter-free Monte-Carlo Tree Search (MTCS), where the space of transcripts is efficiently represented by a tree. We first construct the tree, and then identify its optimal path.",
+ "bbox": [
+ 496,
+ 90,
+ 890,
+ 242
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "4.1. MTCS for New Transcript Generation",
+ "text_level": 1,
+ "bbox": [
+ 498,
+ 252,
+ 830,
+ 268
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "In the tree of transcripts, the root represents the dummy \"start\" of action sequences. The root's descendants sequentially add action classes to the transcripts until leaf nodes, which represent the dummy \"end\". A node $v$ represents the last action class of the path $\\pi_v$ from the root to $v$ , $\\pi_v = \\{\\text{\"start\", } y_1, \\dots, y_v}\\}$ , where $y_v \\in \\mathbb{Y}$ or $y_v = \\text{\"end\"}$ if $v$ is a leaf node. Each node is assigned a weight, $w(v)$ , specified as the joint likelihood of the corresponding path:",
+ "bbox": [
+ 496,
+ 276,
+ 890,
+ 397
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nw (v) = p \\left(\\left| \\pi_ {v} \\right|\\right) \\prod_ {\\left(u, u ^ {\\prime}\\right) \\in \\pi_ {v}} p \\left(y _ {u ^ {\\prime}} \\mid y _ {u}\\right), \\tag {1}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 576,
+ 409,
+ 890,
+ 444
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "where $p(|\\pi_v|)$ is a prior of the transcript length, $u'$ is a child of $u$ along $\\pi_v$ , and $p(y_{u'} | y_u)$ denotes the transition probability of consecutive actions. In the special case, $p(y_{u'} | \"start\")$ and $p(\"end\" | y_u)$ represent the priors that the transcript begins and ends with classes $y_{u'}, y_u \\in \\mathbb{Y}$ . Both $p(|\\pi_v|)$ and $p(y_{u'} | y_u)$ are estimated from $\\mathcal{D}$ . For $p(|\\pi_v|)$ we learn the Poisson distribution, and for $p(y_{u'} | y_u)$ we estimate the frequency of the corresponding class transitions in $\\mathcal{D}$ .",
+ "bbox": [
+ 496,
+ 455,
+ 890,
+ 575
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "The tree is constructed iteratively using the well-known UCT algorithm [20] which balances a trade-off between exploitation and exploration. In each iteration, the root is sequentially expanded with a path of optimal descendants until the dummy \"end\" leaf or the maximum tree depth is reached. We do not allow illegal expansions, i.e., a path is guaranteed to consist of class transitions seen in $\\mathcal{D}$ . Suppose a path has reached node $u$ without meeting the stopping criterion. Then, UCT adds to the path the optimal node $u'$ from a subset of children, $\\mathrm{ch}(u) \\subset \\mathrm{Ch}(u) = \\{u': p(y_{u'} | y_u) > 0\\}$ :",
+ "bbox": [
+ 496,
+ 577,
+ 890,
+ 729
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nu ^ {\\prime} = \\arg \\max _ {v \\in \\operatorname {c h} (u)} \\left[ w (v) + c \\sqrt {\\frac {2 \\log n (u)}{n (v)}} \\right] \\tag {2}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 557,
+ 739,
+ 890,
+ 781
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "where $\\operatorname{ch}(u)$ is randomly sampled $50\\%$ of $\\operatorname{Ch}(u)$ to enable exploring alternative paths; $w(v)$ is given by (1); $c = \\frac{1}{\\sqrt{2}}$ is the exploration-exploitation trade-off parameter; and $n(v)$ is the number of paths in the current tree that include node $v$ . As the tree iteratively grows, the value of $n(v)$ keeps changing for every node, which enables exploring less visited nodes in the space of transcripts even if $w(v)$ is small.",
+ "bbox": [
+ 496,
+ 791,
+ 890,
+ 900
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "page_number",
+ "text": "13507",
+ "bbox": [
+ 480,
+ 944,
+ 517,
+ 955
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "image",
+ "img_path": "images/a14584df1770b75c98816c93bbaa46f6a2becf24c61097e5214b864531491a21.jpg",
+ "image_caption": [
+ "Figure 2. Our MVGA consists of four steps. 1-An action segmentation model is pre-trained using the original training set. 2-New transcripts are generated by selecting optimal paths in a UCT tree [20] of legal transcripts. For every new transcript, a new video is constructed by copying instances of action classes in the new transcript from real videos. 3- Features of the original and constructed new videos are augmented with the Actor and Critic networks. The Actor predicts the amount and location of feature modifications in the video. The Critic estimates the expected rewards for the Actor's two predictions. Both Actor and Critic are learned using the pre-trained action segmenter as oracle. 4- The action segmenter is fine-tuned on the original and augmented training videos."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 163,
+ 88,
+ 810,
+ 338
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "After 1000 tree-growing iterations, the node $v^{*}$ with the highest likelihood $v^{*} = \\arg \\max_{v} w(v)$ in the tree uniquely identifies the newly generated transcript $\\pi_{v^{*}}$ . For generating another transcript, we construct another tree anew.",
+ "bbox": [
+ 75,
+ 452,
+ 470,
+ 513
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "4.2. Generating New Videos of New Transcripts",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 521,
+ 444,
+ 537
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Given a new transcript, $\\pi$ , we sequentially construct a new video with deep features (e.g., I3D [5]) following the ordering of action classes in $\\pi$ . It is worth emphasizing that our video generation occurs directly in the deep feature space. We begin by selecting a video $\\mathcal{X}_0 \\in \\mathcal{D}$ whose ground-truth transcript has the smallest edit distance to $\\pi$ . From $\\mathcal{X}_0$ , we remove all action segments that are not represented in $\\pi$ , resulting in our initial new video $\\mathcal{X}_0'$ . For action classes that are present in $\\pi$ but missing in $\\mathcal{X}_0'$ , we identify the second closest video $\\mathcal{X}_1 \\in \\mathcal{D}$ with these missing actions, and copy their respective temporal intervals to the appropriate locations in the new video, resulting in $\\mathcal{X}_1'$ . This is repeated until the entire $\\pi$ is fully represented by $\\mathcal{X}_g'$ , $g = 0, 1, 2, \\ldots$ . Note that $\\mathcal{X}_g'$ keeps the original lengths of action instances found in real videos, which ensures temporal coherence of every action in $\\mathcal{X}_g'$ . A pseudo-code of the proposed video generation is given in the supplement.",
+ "bbox": [
+ 75,
+ 545,
+ 470,
+ 803
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "5. Augmentation of Frame Features",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 815,
+ 380,
+ 832
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Features of both original and new videos are augmented by a Markov Game (MG) [15, 30, 37]. Since MG is a well-studied framework and we do not claim novelty in our particular formulation, below we rely that the reader is al-",
+ "bbox": [
+ 75,
+ 839,
+ 470,
+ 902
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "ready familiar with the motivation and main concepts of MG. Our MG consists of two agents which sequentially take independent actions causing state changes of a fully observable environment. The environment at step $k$ is defined by state $s_k = (\\mathcal{X}_k, \\mathcal{M}_k)$ , where $\\mathcal{X}_k = \\{x_{k,t} : t = 1, \\dots, T\\}$ is the set of current video frame features, and $\\mathcal{M}_k = \\{m_{k,t} : t = 1, \\dots, T\\}$ is a binary mask assigned to the video frames, $m_{k,t} \\in \\{0,1\\}$ , for keeping the record which frames have been already modified. Given $s_k$ , the two agents follow their respective policies, $\\mu^1$ and $\\mu^2$ , to take actions $a_k^1 = \\mu^1(s_k) \\in \\mathcal{A}^1$ and $a_k^2 = \\mu^2(s_k) \\in \\mathcal{A}^2$ . $\\mathcal{A}^1$ is a continuous action space of agent 1, where $a_k^1 = \\{a_{t,k}^1 : t = 1, \\dots, T\\}$ represents the amount of feature augmentation, i.e., offset features. $\\mathcal{A}^2$ is a discrete action space of agent 2 for selecting video frames for augmentation, $a_k^2 \\in \\{0,1\\}^T$ , where $a_{t,k}^2 = 0$ means that the frame $t$ will not be augmented in step $k$ . $a_k^1$ and $a_k^2$ cause the environment to change to next state $s_{k+1}$ , which incurs the respective two rewards $R_k^1 = R^1(s_k, s_{k+1}, f_\\theta)$ and $R_k^2 = R^2(s_k, a_2^k)$ . Our goal is to learn $\\mu^1$ and $\\mu^2$ so as to maximize the agents' action-value functions given by:",
+ "bbox": [
+ 496,
+ 452,
+ 893,
+ 773
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nQ ^ {i} (s, a) = \\mathbb {E} \\left[ \\sum_ {k \\geq 0} (\\gamma) ^ {k} R _ {k} ^ {i} \\mid \\mu^ {i}, a _ {0} ^ {i} = a, s _ {0} = s \\right], i = \\{1, 2 \\} \\tag {3}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 514,
+ 787,
+ 890,
+ 835
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "where $\\mathbb{E}[\\cdot ]$ denotes expected value, $\\gamma = 0.99$ is the discount factor raised to the power of $k$ , $s\\in S$ , and $a\\in \\mathcal{A}^i$ .",
+ "bbox": [
+ 496,
+ 837,
+ 890,
+ 867
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "For a given video, MG starts from the initial state $s_0 = (\\mathcal{X}_0, \\mathcal{M}_0)$ , with the original frame features and all-zero",
+ "bbox": [
+ 500,
+ 869,
+ 893,
+ 901
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "page_number",
+ "text": "13508",
+ "bbox": [
+ 480,
+ 944,
+ 517,
+ 955
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "mask, $\\mathcal{X}_0 = \\mathcal{X}$ and $\\mathcal{M}_0 = \\{0\\}^T$ . In state $s_k$ , the two agents take their respective actions, which gives $s_{k + 1}$ specified as",
+ "bbox": [
+ 76,
+ 90,
+ 468,
+ 122
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathcal {X} _ {k + 1} = \\left\\{x _ {t, k} + a _ {t, k} ^ {2} \\left(a _ {t, k} ^ {1} + \\epsilon\\right): t = 1, \\dots , T \\right\\}, \\tag {4}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 109,
+ 131,
+ 467,
+ 148
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathcal {M} _ {k + 1} = \\mathcal {M} _ {k} \\vee a _ {k} ^ {2}, \\tag {5}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 101,
+ 152,
+ 468,
+ 170
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "where $\\vee$ is the logical OR operator, and $\\epsilon \\in \\mathbb{R}^{d_{\\mathrm{in}}}$ is noise sampled from the zero-mean and unit-variance Gaussian distribution. $\\epsilon$ enables an exploration of $\\mathcal{A}^1$ and provides a way to generate multiple, distinct, augmented features from the same initial state $s_0$ . From (4), only frames selected by $a_k^2$ get updated with the corresponding offset features $(a_k^1 + \\epsilon)$ . Also, from (5), mask $\\mathcal{M}_k$ keeps the record of previously selected frames for augmentation.",
+ "bbox": [
+ 76,
+ 179,
+ 468,
+ 301
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "MG stops as soon as one of the following happens (experimentally optimized): after $k = 10$ steps, or when $95\\%$ of the video has been augmented, $\\sum_{t=1}^{T} m_{t,k} \\geq 0.95T$ .",
+ "bbox": [
+ 76,
+ 301,
+ 468,
+ 348
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "5.1. Two Rewards for Feature Augmentation",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 354,
+ 421,
+ 371
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "The policy of agent 1, $\\mu^1$ , is learned to augment original features, such that they become more challenging for the pre-trained action segmenter, $f_{\\theta}$ , and thus provide for a more robust subsequent training of $f_{\\theta}$ on the augmented training dataset. This is enforced by specifying the following reward $R_k^1 = R^1(s_k, s_{k+1}, f_{\\theta})$ . Let $\\hat{y}_t$ and $\\hat{y}_t'$ denote the top two scoring class predictions for $x_{t,k}$ by $f_{\\theta}$ , $\\hat{y}_t = \\arg \\max_{y \\in \\mathbb{Y}} p(y|x_{t,k};\\theta)$ and $\\hat{y}_t' = \\arg \\max_{y \\in \\mathbb{Y} \\setminus \\{\\hat{y}_t\\}} p(y|x_{t,k};\\theta)$ . We penalize agent 1 with a negative reward whenever $a_k^1$ causes $f_{\\theta}$ to make a wrong prediction, $\\hat{y}_t \\neq y_t$ . We assign a positive reward to agent 1 when $f_{\\theta}$ 's prediction is equal to the ground truth, $\\hat{y}_t = y_t$ , and reduce this positive reward if the feature augmentation is not challenging enough for $f_{\\theta}$ . The positive reward reduction is proportional to $f_{\\theta}$ 's confidence in its prediction, specified as a difference between its top two scoring predictions, $\\kappa = [p(\\hat{y}_t|x_{t,k};\\theta) - p(\\hat{y}_t'|x_{j,k};\\theta)]$ , $0 < \\kappa \\leq 1$ . Hence, when confidence $\\kappa$ is large and close to 1, the positive reward for agent 1 is maximally reduced as",
+ "bbox": [
+ 76,
+ 378,
+ 468,
+ 667
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nR _ {k} ^ {1} = \\frac {1}{\\| a _ {k} ^ {2} \\|} \\sum_ {t \\in \\mathcal {T} (a _ {k} ^ {2})} r ^ {1} (x _ {t, k + 1}, f _ {\\theta}),\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 135,
+ 680,
+ 367,
+ 719
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nr ^ {1} \\left(x _ {t, k}, f _ {\\theta}\\right) = \\left\\{ \\begin{array}{l l} 1 - \\left[ p \\left(\\hat {y} _ {t} \\mid x _ {t, k}; \\theta\\right) - p \\left(\\hat {y} _ {t} ^ {\\prime} \\mid x _ {t, k}; \\theta\\right) \\right], & \\text {i f} \\hat {y} _ {t} = y _ {t} \\\\ - 1 & , \\text {i f} \\hat {y} _ {t} \\neq y _ {t} \\end{array} \\right. \\tag {6}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 78,
+ 719,
+ 470,
+ 792
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "where $\\mathcal{T}(a_k^2)$ returns the set of frames where $a_{t,k}^2 = 1$ .",
+ "bbox": [
+ 76,
+ 792,
+ 426,
+ 809
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "For the policy of agent 2, $\\mu^2$ , we specify the following three requirements: (i) $\\mu^2$ should not select frames which have already been augmented in the previous MG steps; (ii) frame selection should be local at each MG step and focus only on a very few action instances in the video (and in this way make learning of $\\mu^1$ easier); and (iii) selected temporal",
+ "bbox": [
+ 75,
+ 810,
+ 468,
+ 901
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "intervals should maximally overlap the ground-truth action instances. These three requirements are enforced by specifying the following reward:",
+ "bbox": [
+ 498,
+ 90,
+ 890,
+ 137
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nR _ {k} ^ {2} = \\left\\{ \\begin{array}{c} \\sum_ {\\tau} \\frac {| \\tau \\cap a _ {k} ^ {2} |}{| \\tau |}, \\text {i f} \\sum_ {t} a _ {t, k} ^ {2} m _ {t, k} < \\alpha \\text {a n d} \\| a _ {k} ^ {2} \\| < \\frac {T}{2} \\\\ - 1, \\text {o t h e r w i s e} \\end{array} \\right. \\tag {7}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 504,
+ 147,
+ 890,
+ 229
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "$\\tau$ is the ground-truth mask of actions over frames, and enforces that each iteration of the video augmentation covers the entire time interval of an action rather than randomly scattered frames. $\\alpha = 0.4T$ is an experimentally optimized threshold which allows a flexible frame selection of up to $40\\%$ of the past $\\mathcal{M}_k$ . From (7), agent 2 is penalized when selects more than a half of the video for augmentation.",
+ "bbox": [
+ 496,
+ 229,
+ 890,
+ 335
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "5.2. Learning the Policies for Feature Augmentation",
+ "text_level": 1,
+ "bbox": [
+ 498,
+ 345,
+ 890,
+ 362
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "To learn $\\mu^1$ and $\\mu^2$ , we design an Actor-Critic model with a decentralized-actor deep network for computing the agents' actions, and a centralized-critic deep network for efficiently estimating action-value functions $\\hat{Q}^1$ and $\\hat{Q}^2$ , as shown in Fig. 3. The Actor-Critic framework has been demonstrated effective for continuous agent-action spaces and when the agents have individual rewards [17], as in our case. In general, decentralizing the actor network helps increase stationarity of environments with self-interested agents [37]. The centralized critic network is suitable because it allows relevant information from all of the agents to be shared toward estimating each agent's expected reward.",
+ "bbox": [
+ 496,
+ 368,
+ 890,
+ 549
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "As can be seen in Fig. 3, the input to the actor network, $s_k = (\\mathcal{X}_k, \\mathcal{M}_k)$ , is passed through a backbone, specified as a multi-stage temporal convolutional neural network (TCN) [11], in order to estimate the latent (deep) representation of $s_k$ . Then, two distinct 1-stage TCNs take this latent representation as input and compute the frame selection",
+ "bbox": [
+ 498,
+ 550,
+ 890,
+ 641
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "image",
+ "img_path": "images/ac11775c1bb28e9621f1c50e5b0fac05631f9f124203b8fe256f102618230d56.jpg",
+ "image_caption": [
+ "Figure 3. Actor-Critic model for learning policies $\\mu^1$ and $\\mu^2$ consists of several stages of temporal convolution network (TCN). The actor network predicts feature augmentation, and the centralized critic estimates the expected rewards for each policy."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 504,
+ 662,
+ 890,
+ 828
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "page_number",
+ "text": "13509",
+ "bbox": [
+ 480,
+ 944,
+ 517,
+ 955
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "$a_2^k$ and feature offsets of the selected frames $a_k^1$ . Finally, the predicted $a_k^1$ , $a_2^k$ , and latent feature of $s_k$ are passed to the centralized critic - specifically, a 2-stage TCN - to estimate the expected rewards $\\hat{Q}_k^1 = \\hat{Q}^1 (s_k,a_k^1,a_k^2)$ and $\\hat{Q}_k^2 = \\hat{Q}^2 (s_k,a_k^1,a_k^2)$ of $\\mu^1$ and $\\mu^2$ .",
+ "bbox": [
+ 76,
+ 90,
+ 467,
+ 169
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Parameters of the proposed actor-critic deep architecture are learned using the standard temporal difference learning [21]. The critic learns to iteratively simulate the action-value function, $\\hat{Q}_k^i$ , $i = \\{1, 2\\}$ , which is later used to update parameters of the actor network. To this end, we minimize the following loss of the critic network:",
+ "bbox": [
+ 76,
+ 169,
+ 467,
+ 258
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nL _ {C} = \\sum_ {i \\in \\{1, 2 \\}} \\left(R _ {k} ^ {i} + \\gamma \\hat {Q} _ {k + 1} ^ {i} - \\hat {Q} _ {k} ^ {i}\\right) ^ {2}, \\tag {8}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 143,
+ 271,
+ 467,
+ 310
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "where $R_{k}^{i}$ is given by (6) and (7), and $\\gamma$ is the discount factor. The actor network is learned with the following loss:",
+ "bbox": [
+ 76,
+ 324,
+ 467,
+ 354
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nL _ {A} ^ {i} = \\left\\| a _ {k} ^ {i} \\right\\| _ {2} ^ {2} \\cdot \\left(R _ {k} ^ {i} + \\gamma \\hat {Q} _ {k + 1} ^ {i} - \\hat {Q} _ {k} ^ {i}\\right) ^ {2}, i = \\{1, 2 \\}. \\tag {9}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 91,
+ 378,
+ 467,
+ 407
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "A proof in [37] shows that minimizing $L_{C}$ and $L_{A}$ in (8) and (9) optimizes $\\mu^{1}$ and $\\mu^{2}$ so they manage to achieve maximal rewards $R^{1}$ and $R^{2}$ given by (6) and (7), as desired.",
+ "bbox": [
+ 76,
+ 417,
+ 467,
+ 464
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "6. Results",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 479,
+ 161,
+ 494
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Datasets include three benchmarks: Breakfast [22], GTEA [12], and 50Salads [36]. For each dataset, we precompute the standard I3D frame features without any self-supervision, fine-tuning, or data augmentation, as in [11]. Breakfast has 1712 videos with 48 action classes, GTEA shows 7 complex activities, each specified in terms of 11 action classes including the background class. 50Salads consists of 50 videos with 17 action classes. For Breakfast, GTEA, and 50Salads, we perform the standard 4-fold, 4-fold, and 5-fold cross validation, respectively.",
+ "bbox": [
+ 75,
+ 506,
+ 467,
+ 656
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Metrics. Mean-of-Frame (MoF) is the average framework classification accuracy. Edit score counts edit operations to make the predicted and ground-truth sequences equivalent. F1 score counts true positives when a temporal intersection between the predicted and ground-truth segments is $10\\%$ , $25\\%$ and $50\\%$ , denoted as F1@10,25,50.",
+ "bbox": [
+ 75,
+ 657,
+ 467,
+ 747
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Implementation details. The Actor-Critic has the same layers as MS-TCN [11]. One stage consists of 10 convolution layers with an increasing dilation rate and feature maps with size 128. The mask backbone and feature backbone each represents a 1-stage TCN. Their outputs are, first, concatenated, then, passed to a 4-stage MS-TCN, and, finally, input to the agents' heads. Agent 1's head is a 2-stage MS-TCN. Agent 2's head is a 1-stage TCN. The critic network is a 2-stage MS-TCN. The learning rate for the actor-critic training is 0.0001. Our training of the meta-model is run for",
+ "bbox": [
+ 75,
+ 750,
+ 467,
+ 900
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "30 epochs. The NMS threshold is set to 0.5, and the minimum IOU is set to 0. The discount factor $\\gamma = 0.99$ . The policy training on Tesla-V100s for GTEA takes 3 hours. The MCTS algorithm runs 1000 node expansions in a couple of seconds. The maximum tree width is 200. The maximum tree depth is $\\hat{\\lambda} + 2\\sqrt{\\hat{\\lambda}}$ where $\\hat{\\lambda}$ is the average transcript length of the training set. The number of epochs in the fine-tuning on the original and augmented data is 50 epochs for MS-TCN, 100 epochs for ASFormer, and 50 epochs for BCN. The meta-training of the meta-model increases complexity compared to baseline approaches. It is important to separate the time for training an action segmentation model, and the time for our meta-training. Our meta-training can be viewed as a part of the dataset preparation, which takes significantly less time than manually collecting and annotating real videos.",
+ "bbox": [
+ 496,
+ 90,
+ 890,
+ 335
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "6.1. Ablation Studies of MS-TCN on GTEA",
+ "text_level": 1,
+ "bbox": [
+ 498,
+ 347,
+ 834,
+ 361
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "The training dataset size. Tab. 1 reports how MS-TCN performance on GTEA changes as a function of two variables: (i) the number of training transcripts; and (ii) the total number of original and augmented videos – both expressed as the increase factor of the original set. Tab. 1 shows that when the original set of training videos is tripled such that there are $20\\%$ of new transcripts, MS-TCN achieves the best performance for all of the metrics, and this setting is used in the sequel. Too many augmentations using the same input video leads to similar data outputs, and consequently to overfitting in training of an action segmentation model. Note that testing the transcript augmentation as a standalone contribution is inappropriate and reduces performance (second row), since new transcripts need to be realized as new videos. These new videos are Frankenstein-like constructs from multiple real videos, and hence have highly unrealistic action transitions, and incoherent and sometimes impossible human motions. Therefore, the new videos require additional frame-feature augmentation to provide useful data augmentation.",
+ "bbox": [
+ 496,
+ 369,
+ 890,
+ 672
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Alternatives for generating transcripts. Tab. 2 compares two alternative methods for generating transcripts – namely, auto-regression and our UCT – in terms of: (i) edit distance to the ground-truth transcripts, and (ii) expected reward given by (1). The two metrics are averaged over the top $10\\%$ and $20\\%$ of generated transcripts that are the closest to the ground-truth transcripts. The auto-regressive model is a Transformer Network [35] trained with the cross-entropy loss to predict the next action class given an input sequence of action classes. From Tab. 2, our UCT generates transcripts which are more similar to the ground-truth ones than those produced by auto-regression. Also, the average likelihoods for UCT and ground truth are nearly the same.",
+ "bbox": [
+ 496,
+ 674,
+ 890,
+ 869
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "RL vs Non-RL for frame-feature augmentation. Tab. 3 compares the proposed reinforcement learning and",
+ "bbox": [
+ 500,
+ 869,
+ 890,
+ 900
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "page_number",
+ "text": "13510",
+ "bbox": [
+ 480,
+ 944,
+ 517,
+ 955
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "table",
+ "img_path": "images/e12669ff5e0d47930af944168521ff2a337ed9ad9b72b9be0fb85e408a3edd73.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Method (The default training schedule is 2x) | mAP | AP50 | AP75 | APs | APm | API |
| RetinaNet-Res101(Teacher network) | 38.9 | 58.0 | 41.5 | 21.0 | 42.8 | 52.4 |
| RetinaNet-Res18(Student network) | 33.4 | 51.8 | 35.1 | 16.9 | 35.6 | 44.9 |
| FKD [45] | 35.9 (+2.5) | 54.4 | 38.0 | 17.9 | 39.1 | 49.1 |
| FGD [40] | 36.2 (+2.8) | 54.7 | 38.6 | 19.5 | 40.1 | 48.4 |
| MGD [41] | 36.6 (+3.2) | 55.1 | 39.1 | 19.5 | 40.3 | 50.7 |
| ours | 37.3 (+3.9) | 56.1 | 39.9 | 19.0 | 41.0 | 51.6 |
| ours*(with the parameter inheriting scheme [17]) | 37.5 (+4.1) | 56.0 | 39.9 | 19.4 | 41.5 | 51.2 |
| RetinaNet-Res101(Teacher network) | 38.9 | 58.0 | 41.5 | 21.0 | 42.8 | 52.4 |
| RetinaNet-Res50(Student network) | 37.4 | 56.7 | 39.6 | 20.0 | 40.7 | 49.7 |
| GID [7] | 39.1 (+1.7) | 59.0 | 42.3 | 22.8 | 43.1 | 52.3 |
| FGD [40] | 39.6 (+2.2) | 58.5 | 42.5 | 22.9 | 43.7 | 53.6 |
| MGD [41] | 39.5 (+2.1) | 58.5 | 42.2 | 21.3 | 43.6 | 53.4 |
| ours | 39.9 (+2.5) | 59.0 | 42.7 | 22.3 | 43.9 | 53.3 |
| ours*(with the parameter inheriting scheme [17]) | 40.2 (+2.8) | 59.3 | 43.0 | 22.3 | 44.4 | 54.0 |
| FasterRCNN-Res101(Teacher network) | 39.8 | 60.1 | 43.3 | 22.5 | 43.6 | 52.8 |
| FasterRCNN-Res50(Student network) | 38.4 | 59.0 | 42.0 | 21.5 | 42.1 | 50.3 |
| FGFI [34] | 39.3 (+0.9) | 59.8 | 42.9 | 22.5 | 42.3 | 52.2 |
| GID [7] | 40.2 (+1.8) | 60.7 | 43.8 | 22.7 | 44.0 | 53.2 |
| FGD [40] | 40.4 (+2.0) | 60.7 | 44.3 | 22.8 | 44.5 | 53.5 |
| MGD [41] | 40.1 (+1.7) | 60.2 | 43.6 | 22.9 | 44.2 | 53.3 |
| ours | 41.0 (+2.6) | 61.5 | 44.7 | 23.5 | 45.5 | 53.4 |
| ours*(with the parameter inheriting scheme [17]) | 41.1 (+2.7) | 61.5 | 44.8 | 24.4 | 45.1 | 54.3 |
| FCOS-Res101(Teacher network) | 40.8 | 60.0 | 44.0 | 24.2 | 44.3 | 52.4 |
| FCOS-Res50(Student network) | 38.5 | 57.7 | 41.0 | 21.9 | 42.8 | 48.6 |
| GID [7] | 42.0 (+3.5) | 60.4 | 45.5 | 25.6 | 45.8 | 54.2 |
| FGD [40] | 42.1 (+3.6) | - | - | 27.0 | 46.0 | 54.6 |
| MGD [41] | 42.2(+3.7) | 60.9 | 45.3 | 26.6 | 46.3 | 54.7 |
| ours | 42.5 (+4.0) | 61.2 | 46.1 | 26.6 | 46.8 | 54.6 |
| ours*(with the parameter inheriting scheme [17]) | 43.1 (+4.6) | 61.7 | 46.7 | 27.3 | 47.1 | 55.1 |
| Reppoints-ResX101(Teacher network) | 44.2 | 65.5 | 47.8 | 26.2 | 48.4 | 58.5 |
| Reppoints-Res50(Student network) | 38.6 | 59.6 | 41.6 | 22.5 | 42.2 | 50.4 |
| FKD [45] | 40.6 (+2.0) | 61.7 | 43.8 | 23.4 | 44.6 | 53.0 |
| FGD [40] | 41.3 (+2.7) | - | - | 24.5 | 45.2 | 54.0 |
| MGD [41] | 41.8 (+3.2) | 62.8 | 44.8 | 24.2 | 45.8 | 55.6 |
| MGD [41]*(with the parameter inheriting scheme [17]) | 42.3 (+3.7) | 63.3 | 45.4 | 24.4 | 46.2 | 55.9 |
| ours | 42.2 (+3.6) | 63.0 | 45.6 | 24.3 | 46.4 | 55.7 |
| ours*(with the parameter inheriting scheme [17]) | 43.0 (+4.4) | 63.8 | 46.1 | 24.4 | 47.2 | 57.1 |
",
+ "bbox": [
+ 114,
+ 87,
+ 859,
+ 583
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Table 3: Object detection performance of our proposed MKD with various teacher-student pairs on the COCO dataset.",
+ "bbox": [
+ 94,
+ 592,
+ 870,
+ 608
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Moreover, our method can benefit from a longer training schedule. For example, when training for 36 epochs, our MKD can further improve the performance of the student model over $4.0\\mathrm{mAP}$ , which surpasses other methods by a larger margin than that for 12 epochs. This demonstrates that introducing a masked autoencoder into knowledge distillation can help the student avoid early saturation and benefit more from the distillation process.",
+ "bbox": [
+ 75,
+ 633,
+ 470,
+ 755
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Semantic segmentation. Our MKD can be easily transferred to segmentation tasks such as semantic segmentation, and we conduct knowledge distillation experiments to show the generality of the proposed method. We choose PspNet-Res101 [47] as the teacher to distill PspNet-Res18 and DeepLabV3-Res18 [3]. As shown in Tab. 5, our method improves the PspNet-Res18 model with $4.85\\mathrm{mIoU}$ , which greatly fills the gap between the student and the teacher. Also, MKD surpasses MGD and other methods by a signif",
+ "bbox": [
+ 75,
+ 763,
+ 470,
+ 902
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "icant margin, which verifies the effectiveness and generality of our method.",
+ "bbox": [
+ 496,
+ 633,
+ 890,
+ 662
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Image classification. Since the core bottleneck is particularly prominent in fine-grained visual tasks that rely on spatial localization information, we mainly evaluate our methods on them. We also present the result of MKD on the classification task in Tab. 6, which further verifies its effectiveness.",
+ "bbox": [
+ 496,
+ 666,
+ 890,
+ 755
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "4.2. Ablation Study",
+ "text_level": 1,
+ "bbox": [
+ 498,
+ 768,
+ 653,
+ 787
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Masking ratio. We first examine the influence of different masking ratios on the performance shown in Fig. 7. Large masking ratios benefit more from longer training schedules. For example, a masking ratio of 0.1 achieves the best under a 1x schedule, while the best one changes to 0.3 under a 3x schedule. This is because a larger masking ratio introduces more difficulty to the distillation process and needs more",
+ "bbox": [
+ 496,
+ 794,
+ 890,
+ 900
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "page_number",
+ "text": "6390",
+ "bbox": [
+ 482,
+ 944,
+ 514,
+ 955
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "table",
+ "img_path": "images/bb32b0653227e6bb7a68244adb51e4bd87681786bec9401e105ba4869ded1945.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Method (The default training schedule is 2x) | mAP | AP50 | AP75 | APs | APm | API |
| RetinaNet-Res101(Teacher network) | 38.9 | 58.0 | 41.5 | 21.0 | 42.8 | 52.4 |
| RetinaNet-Res18(Student network) | 33.4 | 51.8 | 35.1 | 16.9 | 35.6 | 44.9 |
| FKD [45] | 35.9 (+2.5) | 54.4 | 38.0 | 17.9 | 39.1 | 49.1 |
| FGD [40] | 36.2 (+2.8) | 54.7 | 38.6 | 19.5 | 40.1 | 48.4 |
| MGD [41] | 36.6 (+3.2) | 55.1 | 39.1 | 19.5 | 40.3 | 50.7 |
| ours | 37.3 (+3.9) | 56.1 | 39.9 | 19.0 | 41.0 | 51.6 |
| ours*(with the parameter inheriting scheme [17]) | 37.5 (+4.1) | 56.0 | 39.9 | 19.4 | 41.5 | 51.2 |
| RetinaNet-Res101(Teacher network) | 38.9 | 58.0 | 41.5 | 21.0 | 42.8 | 52.4 |
| RetinaNet-Res50(Student network) | 37.4 | 56.7 | 39.6 | 20.0 | 40.7 | 49.7 |
| GID [7] | 39.1 (+1.7) | 59.0 | 42.3 | 22.8 | 43.1 | 52.3 |
| FGD [40] | 39.6 (+2.2) | 58.5 | 42.5 | 22.9 | 43.7 | 53.6 |
| MGD [41] | 39.5 (+2.1) | 58.5 | 42.2 | 21.3 | 43.6 | 53.4 |
| ours | 39.9 (+2.5) | 59.0 | 42.7 | 22.3 | 43.9 | 53.3 |
| ours*(with the parameter inheriting scheme [17]) | 40.2 (+2.8) | 59.3 | 43.0 | 22.3 | 44.4 | 54.0 |
| FasterRCNN-Res101(Teacher network) | 39.8 | 60.1 | 43.3 | 22.5 | 43.6 | 52.8 |
| FasterRCNN-Res50(Student network) | 38.4 | 59.0 | 42.0 | 21.5 | 42.1 | 50.3 |
| FGFI [34] | 39.3 (+0.9) | 59.8 | 42.9 | 22.5 | 42.3 | 52.2 |
| GID [7] | 40.2 (+1.8) | 60.7 | 43.8 | 22.7 | 44.0 | 53.2 |
| FGD [40] | 40.4 (+2.0) | 60.7 | 44.3 | 22.8 | 44.5 | 53.5 |
| MGD [41] | 40.1 (+1.7) | 60.2 | 43.6 | 22.9 | 44.2 | 53.3 |
| ours | 41.0 (+2.6) | 61.5 | 44.7 | 23.5 | 45.5 | 53.4 |
| ours*(with the parameter inheriting scheme [17]) | 41.1 (+2.7) | 61.5 | 44.8 | 24.4 | 45.1 | 54.3 |
| FCOS-Res101(Teacher network) | 40.8 | 60.0 | 44.0 | 24.2 | 44.3 | 52.4 |
| FCOS-Res50(Student network) | 38.5 | 57.7 | 41.0 | 21.9 | 42.8 | 48.6 |
| GID [7] | 42.0 (+3.5) | 60.4 | 45.5 | 25.6 | 45.8 | 54.2 |
| FGD [40] | 42.1 (+3.6) | - | - | 27.0 | 46.0 | 54.6 |
| MGD [41] | 42.2(+3.7) | 60.9 | 45.3 | 26.6 | 46.3 | 54.7 |
| ours | 42.5 (+4.0) | 61.2 | 46.1 | 26.6 | 46.8 | 54.6 |
| ours*(with the parameter inheriting scheme [17]) | 43.1 (+4.6) | 61.7 | 46.7 | 27.3 | 47.1 | 55.1 |
| Reppoints-ResX101(Teacher network) | 44.2 | 65.5 | 47.8 | 26.2 | 48.4 | 58.5 |
| Reppoints-Res50(Student network) | 38.6 | 59.6 | 41.6 | 22.5 | 42.2 | 50.4 |
| FKD [45] | 40.6 (+2.0) | 61.7 | 43.8 | 23.4 | 44.6 | 53.0 |
| FGD [40] | 41.3 (+2.7) | - | - | 24.5 | 45.2 | 54.0 |
| MGD [41] | 41.8 (+3.2) | 62.8 | 44.8 | 24.2 | 45.8 | 55.6 |
| MGD [41]*(with the parameter inheriting scheme [17]) | 42.3 (+3.7) | 63.3 | 45.4 | 24.4 | 46.2 | 55.9 |
| ours | 42.2 (+3.6) | 63.0 | 45.6 | 24.3 | 46.4 | 55.7 |
| ours*(with the parameter inheriting scheme [17]) | 43.0 (+4.4) | 63.8 | 46.1 | 24.4 | 47.2 | 57.1 |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.096,
+ 0.593,
+ 0.872,
+ 0.609
+ ],
+ "angle": 0,
+ "content": "Table 3: Object detection performance of our proposed MKD with various teacher-student pairs on the COCO dataset."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.635,
+ 0.471,
+ 0.756
+ ],
+ "angle": 0,
+ "content": "Moreover, our method can benefit from a longer training schedule. For example, when training for 36 epochs, our MKD can further improve the performance of the student model over \\(4.0\\mathrm{mAP}\\), which surpasses other methods by a larger margin than that for 12 epochs. This demonstrates that introducing a masked autoencoder into knowledge distillation can help the student avoid early saturation and benefit more from the distillation process."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.765,
+ 0.471,
+ 0.903
+ ],
+ "angle": 0,
+ "content": "Semantic segmentation. Our MKD can be easily transferred to segmentation tasks such as semantic segmentation, and we conduct knowledge distillation experiments to show the generality of the proposed method. We choose PspNet-Res101 [47] as the teacher to distill PspNet-Res18 and DeepLabV3-Res18 [3]. As shown in Tab. 5, our method improves the PspNet-Res18 model with \\(4.85\\mathrm{mIoU}\\), which greatly fills the gap between the student and the teacher. Also, MKD surpasses MGD and other methods by a signif"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.635,
+ 0.892,
+ 0.664
+ ],
+ "angle": 0,
+ "content": "icant margin, which verifies the effectiveness and generality of our method."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.667,
+ 0.892,
+ 0.756
+ ],
+ "angle": 0,
+ "content": "Image classification. Since the core bottleneck is particularly prominent in fine-grained visual tasks that rely on spatial localization information, we mainly evaluate our methods on them. We also present the result of MKD on the classification task in Tab. 6, which further verifies its effectiveness."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.499,
+ 0.77,
+ 0.655,
+ 0.788
+ ],
+ "angle": 0,
+ "content": "4.2. Ablation Study"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.795,
+ 0.892,
+ 0.901
+ ],
+ "angle": 0,
+ "content": "Masking ratio. We first examine the influence of different masking ratios on the performance shown in Fig. 7. Large masking ratios benefit more from longer training schedules. For example, a masking ratio of 0.1 achieves the best under a 1x schedule, while the best one changes to 0.3 under a 3x schedule. This is because a larger masking ratio introduces more difficulty to the distillation process and needs more"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.483,
+ 0.945,
+ 0.516,
+ 0.957
+ ],
+ "angle": 0,
+ "content": "6390"
+ }
+ ],
+ [
+ {
+ "type": "table",
+ "bbox": [
+ 0.104,
+ 0.102,
+ 0.29,
+ 0.17
+ ],
+ "angle": 0,
+ "content": "| Method (The default training schedule is 2x) | mAP | AP50 | AP75 | APs | APm | API |
| RetinaNet-Res101(Teacher network) | 38.9 | 58.0 | 41.5 | 21.0 | 42.8 | 52.4 |
| RetinaNet-Res18(Student network) | 33.4 | 51.8 | 35.1 | 16.9 | 35.6 | 44.9 |
| FKD [45] | 35.9 (+2.5) | 54.4 | 38.0 | 17.9 | 39.1 | 49.1 |
| FGD [40] | 36.2 (+2.8) | 54.7 | 38.6 | 19.5 | 40.1 | 48.4 |
| MGD [41] | 36.6 (+3.2) | 55.1 | 39.1 | 19.5 | 40.3 | 50.7 |
| ours | 37.3 (+3.9) | 56.1 | 39.9 | 19.0 | 41.0 | 51.6 |
| ours*(with the parameter inheriting scheme [17]) | 37.5 (+4.1) | 56.0 | 39.9 | 19.4 | 41.5 | 51.2 |
| RetinaNet-Res101(Teacher network) | 38.9 | 58.0 | 41.5 | 21.0 | 42.8 | 52.4 |
| RetinaNet-Res50(Student network) | 37.4 | 56.7 | 39.6 | 20.0 | 40.7 | 49.7 |
| GID [7] | 39.1 (+1.7) | 59.0 | 42.3 | 22.8 | 43.1 | 52.3 |
| FGD [40] | 39.6 (+2.2) | 58.5 | 42.5 | 22.9 | 43.7 | 53.6 |
| MGD [41] | 39.5 (+2.1) | 58.5 | 42.2 | 21.3 | 43.6 | 53.4 |
| ours | 39.9 (+2.5) | 59.0 | 42.7 | 22.3 | 43.9 | 53.3 |
| ours*(with the parameter inheriting scheme [17]) | 40.2 (+2.8) | 59.3 | 43.0 | 22.3 | 44.4 | 54.0 |
| FasterRCNN-Res101(Teacher network) | 39.8 | 60.1 | 43.3 | 22.5 | 43.6 | 52.8 |
| FasterRCNN-Res50(Student network) | 38.4 | 59.0 | 42.0 | 21.5 | 42.1 | 50.3 |
| FGFI [34] | 39.3 (+0.9) | 59.8 | 42.9 | 22.5 | 42.3 | 52.2 |
| GID [7] | 40.2 (+1.8) | 60.7 | 43.8 | 22.7 | 44.0 | 53.2 |
| FGD [40] | 40.4 (+2.0) | 60.7 | 44.3 | 22.8 | 44.5 | 53.5 |
| MGD [41] | 40.1 (+1.7) | 60.2 | 43.6 | 22.9 | 44.2 | 53.3 |
| ours | 41.0 (+2.6) | 61.5 | 44.7 | 23.5 | 45.5 | 53.4 |
| ours*(with the parameter inheriting scheme [17]) | 41.1 (+2.7) | 61.5 | 44.8 | 24.4 | 45.1 | 54.3 |
| FCOS-Res101(Teacher network) | 40.8 | 60.0 | 44.0 | 24.2 | 44.3 | 52.4 |
| FCOS-Res50(Student network) | 38.5 | 57.7 | 41.0 | 21.9 | 42.8 | 48.6 |
| GID [7] | 42.0 (+3.5) | 60.4 | 45.5 | 25.6 | 45.8 | 54.2 |
| FGD [40] | 42.1 (+3.6) | - | - | 27.0 | 46.0 | 54.6 |
| MGD [41] | 42.2(+3.7) | 60.9 | 45.3 | 26.6 | 46.3 | 54.7 |
| ours | 42.5 (+4.0) | 61.2 | 46.1 | 26.6 | 46.8 | 54.6 |
| ours*(with the parameter inheriting scheme [17]) | 43.1 (+4.6) | 61.7 | 46.7 | 27.3 | 47.1 | 55.1 |
| Reppoints-ResX101(Teacher network) | 44.2 | 65.5 | 47.8 | 26.2 | 48.4 | 58.5 |
| Reppoints-Res50(Student network) | 38.6 | 59.6 | 41.6 | 22.5 | 42.2 | 50.4 |
| FKD [45] | 40.6 (+2.0) | 61.7 | 43.8 | 23.4 | 44.6 | 53.0 |
| FGD [40] | 41.3 (+2.7) | - | - | 24.5 | 45.2 | 54.0 |
| MGD [41] | 41.8 (+3.2) | 62.8 | 44.8 | 24.2 | 45.8 | 55.6 |
| MGD [41]*(with the parameter inheriting scheme [17]) | 42.3 (+3.7) | 63.3 | 45.4 | 24.4 | 46.2 | 55.9 |
| ours | 42.2 (+3.6) | 63.0 | 45.6 | 24.3 | 46.4 | 55.7 |
| ours*(with the parameter inheriting scheme [17]) | 43.0 (+4.4) | 63.8 | 46.1 | 24.4 | 47.2 | 57.1 |
+
+Table 3: Object detection performance of our proposed MKD with various teacher-student pairs on the COCO dataset.
+
+Moreover, our method can benefit from a longer training schedule. For example, when training for 36 epochs, our MKD can further improve the performance of the student model over $4.0\mathrm{mAP}$ , which surpasses other methods by a larger margin than that for 12 epochs. This demonstrates that introducing a masked autoencoder into knowledge distillation can help the student avoid early saturation and benefit more from the distillation process.
+
+Semantic segmentation. Our MKD can be easily transferred to segmentation tasks such as semantic segmentation, and we conduct knowledge distillation experiments to show the generality of the proposed method. We choose PspNet-Res101 [47] as the teacher to distill PspNet-Res18 and DeepLabV3-Res18 [3]. As shown in Tab. 5, our method improves the PspNet-Res18 model with $4.85\mathrm{mIoU}$ , which greatly fills the gap between the student and the teacher. Also, MKD surpasses MGD and other methods by a signif
+
+icant margin, which verifies the effectiveness and generality of our method.
+
+Image classification. Since the core bottleneck is particularly prominent in fine-grained visual tasks that rely on spatial localization information, we mainly evaluate our methods on them. We also present the result of MKD on the classification task in Tab. 6, which further verifies its effectiveness.
+
+# 4.2. Ablation Study
+
+Masking ratio. We first examine the influence of different masking ratios on the performance shown in Fig. 7. Large masking ratios benefit more from longer training schedules. For example, a masking ratio of 0.1 achieves the best under a 1x schedule, while the best one changes to 0.3 under a 3x schedule. This is because a larger masking ratio introduces more difficulty to the distillation process and needs more
+
+| Method (The default training schedule is 2x) | mAP | AP50 | AP75 | APs | APm | API |
| RetinaNet-Res101(Teacher network) | 38.9 | 58.0 | 41.5 | 21.0 | 42.8 | 52.4 |
| RetinaNet-Res18(Student network) | 33.4 | 51.8 | 35.1 | 16.9 | 35.6 | 44.9 |
| FKD [45] | 35.9 (+2.5) | 54.4 | 38.0 | 17.9 | 39.1 | 49.1 |
| FGD [40] | 36.2 (+2.8) | 54.7 | 38.6 | 19.5 | 40.1 | 48.4 |
| MGD [41] | 36.6 (+3.2) | 55.1 | 39.1 | 19.5 | 40.3 | 50.7 |
| ours | 37.3 (+3.9) | 56.1 | 39.9 | 19.0 | 41.0 | 51.6 |
| ours*(with the parameter inheriting scheme [17]) | 37.5 (+4.1) | 56.0 | 39.9 | 19.4 | 41.5 | 51.2 |
| RetinaNet-Res101(Teacher network) | 38.9 | 58.0 | 41.5 | 21.0 | 42.8 | 52.4 |
| RetinaNet-Res50(Student network) | 37.4 | 56.7 | 39.6 | 20.0 | 40.7 | 49.7 |
| GID [7] | 39.1 (+1.7) | 59.0 | 42.3 | 22.8 | 43.1 | 52.3 |
| FGD [40] | 39.6 (+2.2) | 58.5 | 42.5 | 22.9 | 43.7 | 53.6 |
| MGD [41] | 39.5 (+2.1) | 58.5 | 42.2 | 21.3 | 43.6 | 53.4 |
| ours | 39.9 (+2.5) | 59.0 | 42.7 | 22.3 | 43.9 | 53.3 |
| ours*(with the parameter inheriting scheme [17]) | 40.2 (+2.8) | 59.3 | 43.0 | 22.3 | 44.4 | 54.0 |
| FasterRCNN-Res101(Teacher network) | 39.8 | 60.1 | 43.3 | 22.5 | 43.6 | 52.8 |
| FasterRCNN-Res50(Student network) | 38.4 | 59.0 | 42.0 | 21.5 | 42.1 | 50.3 |
| FGFI [34] | 39.3 (+0.9) | 59.8 | 42.9 | 22.5 | 42.3 | 52.2 |
| GID [7] | 40.2 (+1.8) | 60.7 | 43.8 | 22.7 | 44.0 | 53.2 |
| FGD [40] | 40.4 (+2.0) | 60.7 | 44.3 | 22.8 | 44.5 | 53.5 |
| MGD [41] | 40.1 (+1.7) | 60.2 | 43.6 | 22.9 | 44.2 | 53.3 |
| ours | 41.0 (+2.6) | 61.5 | 44.7 | 23.5 | 45.5 | 53.4 |
| ours*(with the parameter inheriting scheme [17]) | 41.1 (+2.7) | 61.5 | 44.8 | 24.4 | 45.1 | 54.3 |
| FCOS-Res101(Teacher network) | 40.8 | 60.0 | 44.0 | 24.2 | 44.3 | 52.4 |
| FCOS-Res50(Student network) | 38.5 | 57.7 | 41.0 | 21.9 | 42.8 | 48.6 |
| GID [7] | 42.0 (+3.5) | 60.4 | 45.5 | 25.6 | 45.8 | 54.2 |
| FGD [40] | 42.1 (+3.6) | - | - | 27.0 | 46.0 | 54.6 |
| MGD [41] | 42.2(+3.7) | 60.9 | 45.3 | 26.6 | 46.3 | 54.7 |
| ours | 42.5 (+4.0) | 61.2 | 46.1 | 26.6 | 46.8 | 54.6 |
| ours*(with the parameter inheriting scheme [17]) | 43.1 (+4.6) | 61.7 | 46.7 | 27.3 | 47.1 | 55.1 |
| Reppoints-ResX101(Teacher network) | 44.2 | 65.5 | 47.8 | 26.2 | 48.4 | 58.5 |
| Reppoints-Res50(Student network) | 38.6 | 59.6 | 41.6 | 22.5 | 42.2 | 50.4 |
| FKD [45] | 40.6 (+2.0) | 61.7 | 43.8 | 23.4 | 44.6 | 53.0 |
| FGD [40] | 41.3 (+2.7) | - | - | 24.5 | 45.2 | 54.0 |
| MGD [41] | 41.8 (+3.2) | 62.8 | 44.8 | 24.2 | 45.8 | 55.6 |
| MGD [41]*(with the parameter inheriting scheme [17]) | 42.3 (+3.7) | 63.3 | 45.4 | 24.4 | 46.2 | 55.9 |
| ours | 42.2 (+3.6) | 63.0 | 45.6 | 24.3 | 46.4 | 55.7 |
| ours*(with the parameter inheriting scheme [17]) | 43.0 (+4.4) | 63.8 | 46.1 | 24.4 | 47.2 | 57.1 |
",
+ "image_path": "6babc2353137660ba341a8052db2845c87622df466c92074e1ae6047a264a665.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 58,
+ 469,
+ 533,
+ 482
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 58,
+ 469,
+ 533,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 58,
+ 469,
+ 533,
+ 482
+ ],
+ "type": "text",
+ "content": "Table 3: Object detection performance of our proposed MKD with various teacher-student pairs on the COCO dataset."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 46,
+ 502,
+ 288,
+ 598
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 502,
+ 288,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 502,
+ 288,
+ 598
+ ],
+ "type": "text",
+ "content": "Moreover, our method can benefit from a longer training schedule. For example, when training for 36 epochs, our MKD can further improve the performance of the student model over "
+ },
+ {
+ "bbox": [
+ 46,
+ 502,
+ 288,
+ 598
+ ],
+ "type": "inline_equation",
+ "content": "4.0\\mathrm{mAP}"
+ },
+ {
+ "bbox": [
+ 46,
+ 502,
+ 288,
+ 598
+ ],
+ "type": "text",
+ "content": ", which surpasses other methods by a larger margin than that for 12 epochs. This demonstrates that introducing a masked autoencoder into knowledge distillation can help the student avoid early saturation and benefit more from the distillation process."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 46,
+ 605,
+ 288,
+ 715
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 605,
+ 288,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 605,
+ 288,
+ 715
+ ],
+ "type": "text",
+ "content": "Semantic segmentation. Our MKD can be easily transferred to segmentation tasks such as semantic segmentation, and we conduct knowledge distillation experiments to show the generality of the proposed method. We choose PspNet-Res101 [47] as the teacher to distill PspNet-Res18 and DeepLabV3-Res18 [3]. As shown in Tab. 5, our method improves the PspNet-Res18 model with "
+ },
+ {
+ "bbox": [
+ 46,
+ 605,
+ 288,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "4.85\\mathrm{mIoU}"
+ },
+ {
+ "bbox": [
+ 46,
+ 605,
+ 288,
+ 715
+ ],
+ "type": "text",
+ "content": ", which greatly fills the gap between the student and the teacher. Also, MKD surpasses MGD and other methods by a signif"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 304,
+ 502,
+ 545,
+ 525
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 502,
+ 545,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 502,
+ 545,
+ 525
+ ],
+ "type": "text",
+ "content": "icant margin, which verifies the effectiveness and generality of our method."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 304,
+ 528,
+ 545,
+ 598
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 528,
+ 545,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 528,
+ 545,
+ 598
+ ],
+ "type": "text",
+ "content": "Image classification. Since the core bottleneck is particularly prominent in fine-grained visual tasks that rely on spatial localization information, we mainly evaluate our methods on them. We also present the result of MKD on the classification task in Tab. 6, which further verifies its effectiveness."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 305,
+ 609,
+ 400,
+ 624
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 305,
+ 609,
+ 400,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 305,
+ 609,
+ 400,
+ 624
+ ],
+ "type": "text",
+ "content": "4.2. Ablation Study"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 304,
+ 629,
+ 545,
+ 713
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 629,
+ 545,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 629,
+ 545,
+ 713
+ ],
+ "type": "text",
+ "content": "Masking ratio. We first examine the influence of different masking ratios on the performance shown in Fig. 7. Large masking ratios benefit more from longer training schedules. For example, a masking ratio of 0.1 achieves the best under a 1x schedule, while the best one changes to 0.3 under a 3x schedule. This is because a larger masking ratio introduces more difficulty to the distillation process and needs more"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 295,
+ 748,
+ 315,
+ 757
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 295,
+ 748,
+ 315,
+ 757
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 295,
+ 748,
+ 315,
+ 757
+ ],
+ "type": "text",
+ "content": "6390"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 6
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 63,
+ 80,
+ 177,
+ 134
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 63,
+ 80,
+ 177,
+ 134
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 63,
+ 80,
+ 177,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 63,
+ 80,
+ 177,
+ 134
+ ],
+ "type": "table",
+ "html": "