Skip to content

Latest commit

 

History

History
70 lines (52 loc) · 4.94 KB

README.md

File metadata and controls

70 lines (52 loc) · 4.94 KB

REFLECT: Summarizing Robot Experiences for Failure Explanation and Correction

Columbia University in the City of New York

Project Page | arXiV | Summary Video | Dataset


REFLECT is a framework that that leverages Large Language Models (LLMs) for robot failure explanation and correction, based on a summary of robot past experiences generated from multimodal sensory data. The framework is able to tackle both execution and planning failures across a veriety of tasks.

Teaser

Method

1. Hierarchical Summary of Robot Experiences

Summary Hierarchical Robot Summary is composed of: a) a sensory-input summary that converts multi-sensory robot data (RGB-D, sound, robot states) into task-informed scene graphs and audio summary; b) an event-based summary that generates captions for key event frames; c) a subgoal-based summary that contains the end frame of each subgoal.

2. Progressive Failure Explanation with LLM

Reasoning Progressive Failure Explanation verifies success for each subgoal. If a subgoal fails, the algorithm enters the execution analysis stage for detailed explanation on the execution. If all subgoals are satisfied, the algorithm enters planning analysis stage to check errors in the robot plan.

3. Failure Correction Planner

The failure correction planner generates an executable plan for the robot to recover from the failure and complete the task by taking in the task description, the original robot plan, the failure explanation, the final state of the original execution, and the expected goal state.

Codebase

To get started, create a conda environment from the provided env.yml file:

conda env create -f env.yml
conda activate reflect_env
python -m pip install git+https://github.com/openai/CLIP.git

The dependencies have been tested on Ubuntu 20.04.6 LTS, Python 3.9.13. We provide a demo colab that shows how one can generate failure data in the AI2THOR simulation and use the REFLECT framework to analyze and correct the failures.

Issues or pull requests are welcome!

Extensions

Here are few suggestions for extending the framework:

  1. Generate more failure scenarios (more tasks, more scenes, more failure types) in either simulation or real world.
  2. Improve the current perception module by integrating latest vision/audio models that could increase accuracy or capture more low-level information about the environment.
  • [Object detection] You may replace the ground truth object detection obtained from simulation with your own object detector here.
  • [Object state] The codebase supports ground truth object states obtained from simulation and state detection using CLIP. You may replace it with your own method here.
  • [Spatial relations] The codebase uses heuristics on object point clouds to compute inter-object spatial relations. You may replace it with your own method here.
  • [Audio] The codebase uses WAV2CLIP for audio detection. You may replace it with your own model here.
  1. Integrate the framework with an existing policy.
  2. More applications of the robot experience summary, such as safety or efficiency analysis of robot behaviors, or for human-robot collaboration.

Feel free to contact Zeyi at liuzeyi [at] stanford [dot] edu if you have ideas to extend the framework or applying it on new tasks!

Acknowledgement

Some functions of this codebase are adapted from the codebase of Language Models as Zero-Shot Planners by Huang et.al and Socratic Models by Zeng et.al.

Citation

If you find this codebase useful, feel free to cite our work!

@article{liu2023reflect,
  title={REFLECT: Summarizing Robot Experiences for Failure Explanation and Correction},
  author={Liu, Zeyi and Bahety, Arpit and Song, Shuran},
  journal={arXiv preprint arXiv:2306.15724},
  year={2023}
}