Octopus

<p align="center" width="100%"> <img src="https://i.mji.rip/2023/10/13/c4bdc6505f2b3f2304bffb5ea196f5a2.png" width="40%" height="80%"> </p> <div> <div align="center"> <font size=5><strong>Octopus: Embodied Vision-Language Programmer from Environmental Feedback</strong></font> <br> <a href='https://jingkang50.github.io/' target='_blank'>Jingkang Yang^*,1</a>&emsp; <a href='https://github.com/dongyh20/' target='_blank'>Yuhao Dong^*,2,5</a>&emsp; <a href='https://github.com/choiszt/' target='_blank'>Shuai Liu^*,3,5</a>&emsp; <a href='https://brianboli.com/' target='_blank'>Bo Li^*,1</a>&emsp; </br> Ziyue Wang^&dagger;,1</a>&emsp; Chencheng Jiang^&dagger;,4</a>&emsp; Haoran Tan^&dagger;,3</a>&emsp; Jiamu Kang^&dagger;,2</a>&emsp; </br> <a href='https://zhangyuanhan-ai.github.io/' target='_blank'>Yuanhan Zhang¹</a>&emsp; <a href='https://kaiyangzhou.github.io/' target='_blank'>Kaiyang Zhou¹</a>&emsp; <a href='https://liuziwei7.github.io/' target='_blank'>Ziwei Liu^1,5,&#x2709</a> </div> <div align="center"> ¹S-Lab, Nanyang Technological University&emsp; ²Tsinghua University&emsp; <br> ³Beijing University of Posts and Telecommunications&emsp;&emsp; <br> ⁴Xi'an Jiaotong University&emsp; ⁵Shanghai AI Laboratory&emsp; </br> ^* Equal Contribution&emsp; ^&dagger; Equal Engineering Contribution&emsp; ^&#x2709 Corresponding Author </div>

---

Project Page | Octopus Paper | Demo Video

🐙 Introducing Octopus

Octopus is a novel VLM designed to proficiently decipher an agent’s vision and textual task objectives and to formulate intricate action sequences and generate executable code. We provide two models based on the following architectures. Please click

• LLaVA
• Otter

🪸 Introducing OctoVerse

OctoVerse contains three sub-worlds | | OS (tested) | Environment Goal | |---------------|------------------|------------------------------------| | OctoGibson | Ubuntu 20.04 | 500 Tasks on OmniGibson | | OctoGTA | Windows 11 | 20 Tasks to evaluate transfer learning | | OctoMC | Ubuntu/Windows | 20 Tasks to evaluate transfer learning on MineCraft worlds, such as making an axe. |

• Training data collection pipeline in octogibson environment
• Evaluation pipeline in octogibson environment
• Evaluation pipeline in octogta environment
• Training pipeline of the octopus model

Contact: Leave issue or contact jingkang001@e.ntu.edu.sg and dongyh20@mails.tsinghua.edu.cn. We are on call to respond.

🦾 Updates

[2023-10]

1. 🤗 Introducing Project Octopus' homepage: https://choiszt.github.io/Octopus.
2. 🤗 Check our paper introducing Octopus in details.

🏁 Get Started

1. Training Data Collection: For data collection from octogibson environment, you need to set up two conda environments: omnigibson and gpt4. The omnigibson environment has an agent to act following the instruction from gpt4 environment. Please checkout here for detailed information.
2. Evaluation in OctoGibson: We provide the pipeline that the simulator sends messages to the Octopus server and gets responses to control the agent.
3. Evaluation in OctoGTA: We provide instructions, code, and MOD so that the Octopus can complete tasks in the GTA environment. Please checkout here for detailed information.
4. Octopus Training: We provide code for training Octopus. Please checkout here for detailed information.

📑 Citation

If you found this repository useful, please consider citing:

@article{yang2023octopus,
    title = {Octopus: Embodied Vision-Language Programmer from Environmental Feedback},
    author = {Jingkang Yang and Yuhao Dong and Shuai Liu and Bo Li and Ziyue Wang and Chencheng Jiang and Haoran Tan and Jiamu Kang and Yuanhan Zhang and Kaiyang Zhou and Ziwei Liu},
    year = {2023},
}

👨‍🏫 Acknowledgements

We thank the OmniGibson team for their help and great contribution to the open-source community.