Datasets:
Tasks:
Visual Question Answering
Formats:
parquet
Languages:
English
Size:
10K - 100K
ArXiv:
License:
| license: cc-by-4.0 | |
| language: | |
| - en | |
| pretty_name: RoboProcessBench | |
| task_categories: | |
| - visual-question-answering | |
| tags: | |
| - robotics | |
| - embodied-ai | |
| - benchmark | |
| - vision-language-models | |
| - process-understanding | |
| - manipulation | |
| size_categories: | |
| - 10K<n<100K | |
| configs: | |
| - config_name: default | |
| data_files: | |
| - split: train | |
| path: splits/processdata_sft.parquet | |
| - split: eval | |
| path: splits/processdata_eval.parquet | |
| # RoboProcessBench | |
| This benchmark is one core contribution of the paper, **"Robo2Process: Benchmarking and Learning Process-Aware Understanding in Vision-Language Robotic Manipulation"**, which has been accpeted at the 40th Conference on Neural Information Processing Systems (NeurIPS 2026). | |
| ## Dataset Summary | |
| RoboProcessBench is a process-aware understanding benchmark for vision-language robotic manipulation. It evaluates whether VLMs can infer how a manipulation execution unfolds, including phase, contact, motion, bimanual coordination, primitive-local progress, temporal order, outcome, and primitive-level transitions. | |
| This release contains 57,892 QA rows: 48,841 SFT rows and 9,051 evaluation rows across 12 task families and 260 manipulation tasks. The split follows strict episode / recording / scene isolation. The benchmark is derived from GM-100, RH20T, REASSEMBLE, and AIST-Bimanual. | |
| ## Task Families | |
| RoboProcessBench contains 12 diagnostic question families: | |
| - **T1 Phase Recognition:** identify the current coarse process phase. | |
| - **T2 Contact Detection:** determine whether task-relevant contact has occurred. | |
| - **T3 Motion Direction Prediction:** infer the dominant motion direction from short temporal context. | |
| - **T4 Bimanual Coordination State:** identify the current coordination state of two arms. | |
| - **T5 Primitive-local Progress:** estimate progress within the current local manipulation step. | |
| - **T6 Motion State Recognition:** distinguish actively moving from stationary states. | |
| - **T7 Operation Outcome Prediction:** predict eventual success or failure from partial execution evidence. | |
| - **T8 Temporal Ordering:** reconstruct the chronological order of shuffled observations. | |
| - **T9 Temporal Priority Prediction:** decide which of two observations occurred earlier. | |
| - **T10 Current Primitive Recognition:** identify the current low-level primitive. | |
| - **T11 Next Primitive Prediction:** infer the next primitive from local process context. | |
| - **T12 Primitive Chain Restoration:** restore a masked primitive in a local primitive chain. | |
| ## Repository Contents | |
| The release is organized as a compact, reviewable benchmark package: | |
| ```text | |
| RoboProcessBench/ | |
| ├── splits/ # SFT and Eval QA entries, split summaries | |
| │ ├── processdata_sft.jsonl | |
| │ ├── processdata_sft.parquet | |
| │ ├── sft_manifest.jsonl | |
| │ ├── processdata_eval.jsonl | |
| │ ├── processdata_eval.parquet | |
| │ └── eval_manifest.jsonl | |
| │ | |
| ├── metadata/ # statistics, schema, licenses, prompt templates, reconstruction notes | |
| │ ├── task_distribution.csv | |
| │ ├── asset_licenses.csv | |
| │ ├── prompt_templates.md | |
| │ └── reconstruction.md | |
| │ | |
| ├── examples/ # rendered representative benchmark cards | |
| │ └── task_cards/ | |
| │ ├── T1_phase_recognition.png | |
| │ ├── T2_contact_detection.png | |
| │ ├── ... | |
| │ └── T12_primitive_chain_restoration.png | |
| │ | |
| ├── ProcessData-SFT-Qwen/ # LoRA adapter weights and training configuration | |
| │ ├── adapter_config.json | |
| │ ├── adapter_model.safetensors | |
| │ ├── ... | |
| │ └── vocab.json | |
| │ | |
| ├── ProcessData-SFT-Qwen_results/ # predictions and summary of the post-trained model | |
| │ ├── ProcessData-SFT-Qwen_predictions.json | |
| │ └── ProcessData-SFT-Qwen_summary.json | |
| │ | |
| ├── benchmark_card.md # benchmark-level documentation | |
| ├── croissant.json # Croissant core + Responsible AI metadata | |
| └── README.md | |
| ``` | |
| Full upstream videos and full frame dumps are not redistributed in this release. | |
| ## License and Attribution | |
| The original benchmark contributions in RoboProcessBench, including newly authored questions, answer annotations, benchmark organization, split manifests, and metadata, are released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. This license applies only to materials for which the RoboProcessBench authors hold the necessary rights. Meanwhile, RoboProcessBench is constructed using four upstream robotic manipulation datasets: | |
| - GM-100 | |
| - RH20T | |
| - REASSEMBLE | |
| - AIST-Bimanual | |
| The original datasets, including their videos, robot trajectories, sensor measurements, and other third-party materials, remain subject to their respective licenses and terms of use. The CC BY 4.0 license does not replace, override, or sublicense any upstream dataset license. Users who obtain or reconstruct visual observations from the original datasets must comply with the applicable upstream licenses and attribution requirements. See `metadata/asset_licenses.csv` for source-specific licensing information. | |
| ## Citation | |
| If you find our work inspiring or helpful, please use the following format for citation: | |
| ```text | |
| Dayu Xia, Yue Shi, Yao Mu, Huiting Ji, Chaofan Ma, Yingjie Zhou, Jiezhang Cao, Hua Chen, Yang Liu, and Guangtao Zhai. Robo2Process: Benchmarking and Learning Process-Aware Understanding in Vision-Language Robotic Manipulation. arXiv preprint arXiv:2606.13040, 2026. | |
| ``` | |