| # ELLSA: End-to-end Listen, Look, Speak and Act |
|
|
| <div align="center"> |
|
|
| <div> |
| <a href="https://arxiv.org/pdf/2510.16756" target="_blank"> |
| <img src="https://img.shields.io/badge/Paper-arXiv-red.svg" alt="Paper arXiv"> |
| </a> |
| <a href="https://github.com/bytedance/SALMONN/tree/ELLSA" target="_blank"> |
| <img src="https://img.shields.io/badge/GitHub-Code-blue" alt="GitHub Code"> |
| </a> |
| <a href="https://huggingface.co/tsinghua-ee/ELLSA" target="_blank"> |
| <img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-yellow" alt="Hugging Face Models"> |
| </a> |
| <a href="https://huggingface.co/datasets/tsinghua-ee/ELLSA_train_data" target="_blank"> |
| <img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Train%20Dataset-yellow" alt="Train data"> |
| </a> |
| <a href="https://huggingface.co/datasets/tsinghua-ee/ELLSA_test_data" target="_blank"> |
| <img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Test%20Dataset-yellow" alt="Test data"> |
| </a> |
| <img src="https://img.shields.io/badge/License-Apache%202.0-green" alt="License"> |
| </div> |
| |
| </div> |
|
|
| The **first** end-to-end model that unifies **vision, speech, text and action** in a **streaming full-duplex** framework, enabling joint multimodal perception and concurrent generation. |
|
|
| <p align="center"> |
| <img src="docs/imgs/ellsa.png" width="60%" height="60%"> |
| </p> |
| |
| ## π§ͺ Highlights |
| * **Full-Duplex Multimodal Interaction**: unifies *listening*, *looking*, *speaking*, and *acting* in a single end-to-end architecture, enabling simultaneous multimodal perception and generation. |
| * **SA-MoE Architecture for Efficient Multimodal Fusion**: utilizes modality-specific experts with shared *attention* to reduce interference and leverage the capabilities of pretrained models. |
| * **Unique Human-like Capabilities**: supports *speaking-while-acting*, *context-grounded VQA*, *instruction rejection*, and *action barge-in*, enabling more natural interactive intelligence. |
|
|
| ## π§ REPO TODO List |
| - [x] Support for evaluation on speech interaction. |
| - [x] Support for evaluation on LIBERO. |
| - [x] Release the training data. |
| - [x] Support for training. |
| - [ ] Support for evaluation on CALVIN. |
|
|
| ## π Experiments |
|
|
| ### Basic Capabilities |
|
|
| On speech-interaction and robotmanipulation benchmarks, ELLSA matches modality-specific baselines. |
|
|
| ##### Speech Interaction |
|
|
| | Model | Llama Q. S2T | Llama Q. S2S | Web Q. S2T | Web Q. S2S | TriviaQA S2T | TriviaQA S2S | AlpacaEval S2T | AlpacaEval S2S | |
| |--------------|--------------|--------------|------------|------------|--------------|--------------|----------------|----------------| |
| | Moshi | 60.8 | 54.5 | 23.4 | 22.1 | 25.6 | 16.7 | 1.84 | 1.76 | |
| | Freeze-Omni | 74.2 | 56.2 | **40.8** | 27.9 | 45.1 | 28.5 | **3.90** | 2.46 | |
| | **ELLSA** | **74.7** | **70.0** | 39.5 | **36.5** | **45.2** | **41.7** | 3.09 | **2.80** | |
|
|
| ##### Speech-conditioned Robot Manipulation |
|
|
| | Model | SPATIAL | OBJECT | GOAL | LONG | Average | |
| |--------------|--------|--------|-------|-------|---------| |
| | DP* | 78.3% | 92.5% | 68.3% | 50.5% | 72.4% | |
| | Octo | 78.9% | 85.7% | 84.6% | 51.1% | 75.1% | |
| | OpenVLA | 84.9% | 88.4% | 79.2% | 53.7% | 76.5% | |
| | SpatialVLA | 88.2% | 89.9% | 78.6% | 55.5% | 78.1% | |
| | CoT-VLA | 87.5% | 91.6% | 87.6% | 69.0% | 81.1% | |
| | Οβ-FAST | **96.4%** | **96.8%** | **88.6%** | 60.2% | 85.5% | |
| | **ELLSA** | 90.8% | 95.8% | 86.4% | **84.4%** | **89.4%** | |
|
|
| ### Advanced Capabilities |
|
|
| ELLSA can accomplish tasks previously unattainable, such as *dialogue and action turn-taking prediction*, *rejection of defective instructions*, *speaking while acting* and *responding to action barge-ins*. These results highlight the feasibility and significance of full-duplex multimodal interaction as a foundation for more natural and general multimodal interactive intelligence. |
|
|
| <div align="center"> |
| <img src="docs/imgs/example.png" width="90%" alt="WAVE Architecture"/> |
| <br> |
| <em>An example of ELLSAβs advanced capabilities: starting from a spoken instruction, the model executes the action, engages in context-grounded VQA, and supports action barge-in. This instance demonstrates not only ELLSAβs core skills but also its unique advanced capabilities: its MIMO capacity to process multimodal inputs and outputs simultaneously, and its duplex capability to manage complex conversational dynamics such as turn-taking and interruptions.</em> |
| </div> |
|
|
| ## π οΈ Setup |
| Here we provide a conda environment setup for the project. |
| ```shell |
| conda create -n ellsa python=3.10 |
| conda activate ellsa |
| pip install -r requirements.txt |
| ``` |
| > If you run into issues installing `flash-attention` or `kaldifeat`, you can instead use the prebuilt wheels available here: [flash-attn prebuilt wheels](https://github.com/Dao-AILab/flash-attention/releases) and [kaldifeat prebuilt wheels](https://csukuangfj.github.io/kaldifeat/cuda.html). |
|
|
| ## π₯ Training |
|
|
| Before running training, make sure to download all required checkpoints and Data. |
|
|
| | **Model** | **Download** | |
| | :--- | :---: | |
| | **Emu3-vision** | [π€ HuggingFace](https://huggingface.co/BAAI/Emu3-VisionTokenizer) | |
| | **UniVLA-LIBERO** | [π€ huggingface](https://huggingface.co/Yuqi1997/UniVLA/tree/main/UNIVLA_LIBERO_VIDEO_BS192_8K) | |
| | **Llama-3.1-8B-Instruct** | [π€ huggingface](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) | |
| | **CosyVoice2-0.5B** | [π€ huggingface](https://huggingface.co/FunAudioLLM/CosyVoice2-0.5B) | |
| | **ELLSA** | [π€ huggingface](https://huggingface.co/tsinghua-ee/ELLSA) | |
|
|
| | **Data** | **Download** | |
| | :--- | :---: | |
| | **Train Data** | [π€ HuggingFace](https://huggingface.co/datasets/tsinghua-ee/ELLSA_train_data) | |
|
|
| > Only the dialogue transcripts are released. The corresponding speech files need to be synthesized using a text-to-speech model. |
|
|
| ```shell |
| pip install -r requirements-training.txt --no-build-isolation |
| ``` |
|
|
| ### Speech Interaction |
|
|
| Train speech understanding and merge the LoRA weights: |
|
|
| ```shell |
| export ELLSA_BASE_PATH="$PWD" |
| export LLAMA_CKPT_PATH=/path/to/Llama-3.1-8B-Instruct |
| export DATA_SPEECH_PATH=/path/to/speech_annotation.json |
| export SPEECH_ENCODER_PATH=/path/to/SPEAR/iter-500000-avg-4.pt |
| bash scripts/simulator/libero/train_speech_only.sh |
| python scripts/merge_lora.py output/ellsa_speech/checkpoint-20000 |
| ``` |
|
|
| ### Robot manipulation on LIBERO Benchmark |
|
|
| Train the mixed model, then add speech generation: |
|
|
| ```shell |
| export UNIVLA_CKPT_PATH=/path/to/UNIVLA_LIBERO_VIDEO_BS192_8K |
| export COSY_CKPT_PATH=/path/to/CosyVoice2-0.5B |
| export SPEECH_EXPERT_PATH="$PWD/output/ellsa_speech/checkpoint-20000-merged" |
| bash scripts/simulator/libero/train_speech_mix_moe.sh |
| export PT_CKPT_PATH="$PWD/output/ellsa_libero_mix/checkpoint-50000" |
| bash scripts/simulator/libero/train_tts_only.sh |
| ``` |
|
|
| ## π Inference |
|
|
| ### Required Checkpoints and Data |
|
|
| Before running inference, make sure to download all required checkpoints and Data. |
|
|
| | **Model** | **Download** | |
| | :--- | :---: | |
| | **Emu3-vision** | [π€ HuggingFace](https://huggingface.co/BAAI/Emu3-VisionTokenizer) | |
| | **UniVLA-LIBERO** | [π€ huggingface](https://huggingface.co/Yuqi1997/UniVLA/tree/main/UNIVLA_LIBERO_VIDEO_BS192_8K) | |
| | **Llama-3.1-8B-Instruct** | [π€ huggingface](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) | |
| | **CosyVoice2-0.5B** | [π€ huggingface](https://huggingface.co/FunAudioLLM/CosyVoice2-0.5B) | |
| | **ELLSA** | [π€ huggingface](https://huggingface.co/tsinghua-ee/ELLSA) | |
|
|
| | **Data** | **Download** | |
| | :--- | :---: | |
| | **Test Data** | [π€ HuggingFace](https://huggingface.co/datasets/tsinghua-ee/ELLSA_test_data) | |
|
|
| ### Speech Interaction |
|
|
| ```shell |
| cd reference/RoboVLMs |
| bash scripts/run_eval_speech_only.sh ${CKPT_PATH} |
| ``` |
|
|
| ### Robot manipulation on LIBERO Benchmark |
|
|
| Build LIBERO environment and dataset based on the [instruction](docs/libero.md). |
|
|
| ```shell |
| cd reference/RoboVLMs |
| bash scripts/run_eval_libero_contemporary.sh ${CKPT_PATH} |
| ``` |
|
|
|
|
| <section class="section"> |
| <div class="container is-max-desktop"> |
| <h2 class="title is-4">π Code Structure</h2> |
| <pre style="background-color: #f9f9f9; padding: 1.25rem; border-radius: 8px; font-size: 14px; overflow-x: auto;"> |
| <span style="color: #6c757d;">ELLSA/</span> |
| βββ <strong>configs/</strong> <span style="color: #6c757d;"># Model configuration files</span> |
| βββ <strong>models/</strong> <span style="color: #6c757d;"># Tokenizer and diffusion test</span> |
| βββ <strong>train/</strong> <span style="color: #6c757d;"># Training dataset and pipeline</span> |
| βββ <strong>reference/</strong> <span style="color: #6c757d;"># Reference code</span> |
| β βββ <strong>cosyvoice/</strong> <span style="color: #6c757d;"># Speech synthesizer</span> |
| β βββ <strong>Emu3/</strong> <span style="color: #6c757d;"># Base code</span> |
| β βββ <strong>RoboVLMs/</strong> <span style="color: #6c757d;"># Evaluation code</span> |
| β βββ <strong>spear_encoder/</strong> <span style="color: #6c757d;"># Speech encoder</span> |
| βββ <strong>scripts/</strong> <span style="color: #6c757d;"># Shell scripts for training</span> |
| βββ <strong>tools/</strong> <span style="color: #6c757d;"># Data preprocessing tools</span> |
| βββ <strong>README.md</strong> <span style="color: #6c757d;"># Project description and user guide</span> |
| </pre> |
| </div> |
| </section> |
| |
| ## β€οΈ Acknowledgement |
| Our work is built upon the following projects, Thanks for their great open-source work! |
| - [Emu3](https://github.com/baaivision/Emu3) |
| - [RoboVLMs](https://github.com/Robot-VLAs/RoboVLMs) |
| - [OpenVLA](https://github.com/openvla/openvla) |
| - [UniVLA](https://github.com/baaivision/UniVLA) |
|
|
| ## π Citation |
| If you find this project useful, please consider citing our work: |
| ```bibtex |
| @inproceedings{wang2026end, |
| title={End-to-end Listen, Look, Speak and Act}, |
| author={Wang, Siyin and Yu, Wenyi and Chen, Xianzhao and Tian, Xiaohai and Zhang, Jun and Lu, Lu and Zhang, Chao}, |
| journal={Proc. ICLR}, |
| year={2026}, |
| address={Rio de Janeiro} |
| } |
| ``` |
