RoboEXP: Action-Conditioned Scene Graph via Interactive Exploration for Robotic Manipulation

Hanxiao Jiang¹, Binghao Huang¹, Ruihai Wu³, Zhuoran Li⁴, Shubham Garg², Hooshang Nayyeri², Shenlong Wang¹, Yunzhu Li¹

¹University of Illinois Urbana-Champaign, ²Amazon, ³Peking University, ⁴National University of Singapore

Website | Paper | Colab | Video

Overview

This repository contains the official implementation of the RoboEXP system for the interactive exploration task.

teaser_video.mp4

Try it in Colab!

In this notebook, we replay an example exploration process using the saved observations and cached decisions.

Setup

# Here we use cuda-11.7
export CUDA_HOME=/usr/local/cuda-11.7/
# create conda environment
conda create -n roboexp python=3.8
conda activate roboexp

# Install the packages
bash scripts/install_env.sh

# download pretrained models
bash scripts/download_pretrained_models.sh

Interactive Exploration

To run our RoboEXP system, you also need to create the file my_apikey and copy your OpenAI API key into the file.

To calibrate the wrist camera, we provide the hand-eye calibration code in calibrate_wrist.py, and you can download our calibration board.

# Download the calibration board
gdown 1KWYncDGjtGthePC3wzCu9zBW0nZI0RBM
# Download the example calibration result
gdown 1b2yp45eJVyXnOg11OImTER3zMRUcaZRN

Run our RoboEXP system

# Set `visualize=True` to enable visualizations
python interactive_explore.py 

Code Structure Explanation

roboexp is the main library for our RoboEXP system, comprising the following key components:

• roboexp/env: This component provides the basic control API for the robot, camera, and calibration setup. It facilitates obtaining observations from the camera and controlling the robot.
• roboexp/perception: This component offers an API for the perception module, enabling image processing to derive 2D semantic information from images.
• roboexp/memory: This component provides an API for the memory module, allowing instance merging in 3D and the construction of our Action-Conditioned Scene Graph (ACSG).
• roboexp/decision: This component offers an API for the decision module, facilitating decision-making based on observations.
• roboexp/act: This component provides an API for the action module, enabling action planning based on decisions and our memory.

Citation

If you find this repo useful for your research, please consider citing the paper

@article{jiang2024roboexp,
    title={RoboEXP: Action-Conditioned Scene Graph via Interactive Exploration for Robotic Manipulation},
    author={Jiang, Hanxiao and Huang, Binghao and Wu, Ruihai and Li, Zhuoran and Garg, Shubham and Nayyeri, Hooshang and Wang, Shenlong and Li, Yunzhu},
    journal={arXiv preprint arXiv:2402.15487},
    year={2024}
}