FLEX: Framework for Learning Robot-Agnostic Force-based Skills Involving Sustained Contact Object Manipulation

October 15, 2024 ยท View on GitHub

This repository contains the codebase for FLEX, a framework for learning force-based reinforcement learning (RL) skills that are robot-agnostic and generalizable to various objects with similar joint configurations. Below, you will find instructions on how to install the dependencies and run FLEX. If you have any questions or wish to contribute or connect, please contact the authors.

Table of Contents

System requirements

This code is tested on Ubuntu 20.02 and Ubuntu 22.04, with RTX30 and RTX40 series GPU.

Clone the Repository

Clone the repository and its submodules (FLEX includes a submodule called contact_graspnet for generating grasp proposals). Please ensure that you only clone the main branch; the website branch is for the project website.

git clone --recurse-submodules https://github.com/tufts-ai-robotics-group/FLEX.git --single-branch

Clone the AO-GRASP Repository

Clone our version of AO-GRASP repository into a separate directory. Do not clone their version of contact_graspnet, as it is known to have compatibility issues with newer GPUs (RTX 30 series onwards).

git clone https://github.com/WenchangGaoT/ao-grasp.git

Dependencies

There are two environments to set up to run FLEX: the main FLEX environment and a separate environment for contact_graspnet due to different CUDA versions.

Setting Up the FLEX Environment

  1. Install Conda: We recommend Miniconda. Follow the Miniconda installation instructions.
  2. Create the FLEX Environment: 
    conda env create --name flex --file=environment.yml
  3. Create the contact_graspnet Environment: contact_graspnet relies on a different version of CUDA, so a separate environment is required. 
    cd flex/grasps/aograsp/contact_graspnet/
conda env create --name cgn --file cgn_env_rtx30.yml

    Note: The environment must be named cgn because a script automatically launches this environment when running the grasp module.

Installing the Packages

  1. Install AO-GRASP: Navigate to the root directory of the cloned AO-GRASP package and run:

    conda activate flex
pip install -e .

    You do not need to reinstall AO-GRASP's dependencies, as they are included in the flex environment's requirements.

  2. Install PointNet++ for AO-GRASP:

    conda activate flex
cd aograsp/models/Pointnet2_PyTorch/
pip install -e .
cd pointnet2_ops_lib/
pip install -e .

  3. Install FLEX: Navigate to the root directory of the cloned FLEX package and run:

    conda activate flex
cd flex/
pip install -e .

Fixing Rendering Issues

If you encounter rendering issues with Robosuite (the robot simulator used in this project), set the following environment variable before running the code:

export MUJOCO_GL="osmesa"

To make this change permanent, add the above line to your ~/.bashrc file.

Try It Out!

You can test the trained policies or train a new policy using the commands below:

  • Test the Trained Policies:

    conda activate flex
cd flex/
python scripts/parallel_test.py --render=true --run_id=$(run_id)

  • Train the Policy:

    conda activate flex
cd flex/
python scripts/parallel_train.py

Connect

For any questions regarding the paper or the code, please contact us at: