Code for RACER: Rich Language-Guided Failure Recovery Policies for Imitation Learning

Getting Started

Install RACER

Tested (Recommended) Versions: Ubuntu 20.04, Python 3.8 and CUDA 11.6.
Step 1: We recommend using conda and creating a virtual environment.

conda create --name racer python=3.8.18
conda activate racer

Step 2: Install PyTorch. For more information visit here.

conda install install pytorch==1.13.0 torchvision==0.14.0 torchaudio==0.13.0 pytorch-cuda=11.6 -c pytorch -c nvidia

Step 3: Install PyTorch3D. For more information visit here.

conda install iopath==0.1.9 -c iopath
conda install nvidiacub==1.10.0 -c bottler 
conda install pytorch3d==0.7.5 -c pytorch3d

Step 4: Install CoppeliaSim. Download and unzip CoppeliaSim_Edu_V4_1_0 according to your Ubuntu version. Once you have downloaded CoppeliaSim, add the following to your ~/.bashrc file and source it.

export COPPELIASIM_ROOT=<PATH/TO/COPPELIASIM/DIR>
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$COPPELIASIM_ROOT
export QT_QPA_PLATFORM_PLUGIN_PATH=$COPPELIASIM_ROOT
export DISPLAY=:<DISPLAY_ID>

And remember to start the vncserver use vncserver :<DISPLAY_ID>. DISPLAY_ID is a positive int number.

Step 5: Clone the submodules and install all packages.

cd <PATH_TO_RACER>
git submodule update --init
pip install -e . 
pip install -e libs/PyRep 
pip install -e libs/RLbench 
pip install -e libs/YARR 
pip install -e libs/peract_colab

Prepare Data

Step 1: Download RLbench dataset provided by PerAct. Please download and place them under RACER/racer/data/rlbench/xxx where xxx is either train, test, or val.
You can only download test for evaluation if you want to save some space.
Step 2: Download our language-guided failure-recovery augmentation training data. put it under RACER/racer/data/augmented_rlbench/xxx, where xxx is either train or val.

For more information about our automatic data generation pipeline, please refer to this page.

Data Processing

After set up the language encoder service following this, you can process the data using

python racer/utils/preprocess_data.py

The data will be processed into replay buffers by YARR.

To save data processing time (around 1-2 days), you can also directly download our generated replay buffer data racer_replay_public here without processing by yourself, and put place it under RACER/racer/replay_buffers.
This is used only if you want to train RACER by yourself and not needed if you just want to evaluate.

Training RACER

Training visuomotor policy

To train RACER on all RLBench tasks, use the following command:

./scripts/train_racer.sh

You can change the rich/simple/task yaml files for --exp_cfg_path to train different types of RACER models, and change exp_id to name the trained model directory name.

Training LLaVA model

Please refer to this page

Evaluating RACER

Setup Service

We set a client-server framework for language encoder service and LLaVA model service that RACER/RVT needs, please refer to this page for details.

After the language encoder service is set up, you can test it with

python racer/utils/lang_enc_utils_v2.py --lm-addr <lm service addr>

After the LLaVA service is set up, you can test it with

python racer/evaluation/llava_api/api.py  --vlm-addr <vlm service addr>

Model Checkpoints

Download the official RVT model and place it as racer/runs/rvt_ckpt.
Download our RACER visuomotor policy model and place it as racer/runs/racer-visuomotor-policy-rich.

Evaluate RVT

./scripts/eval_rvt.sh

Evaluate RACER

./scripts/eval_racer.sh

It takes around 5 hours to finish all tasks evaluation. The evaluation results are stored under the model checkpoint directory.

Peak Memory: 19.2GB for langauge model service, 31.7 GB for llava service, 15.5 GB for visuomotor policy model. Using different GPUs or machines to allocate the memory usage is recommended.

Gradio Online Demo

First install pip install gradio==4.36.1, then run

./scripts/demo.sh

More detailed cases can be found here

Acknowledgement

This code is adapted and modified upon the released RVT code. We really appreciate the authors release their code, which is very helpful to our research!

This work is supported in part by NSF IIS1949634, NSF SES-2128623, and has benefited from the Microsoft Accelerate Foundation Models Research (AFMR) grant program.

Citation

@article{dai2024racer,
  title={RACER: Rich Language-Guided Failure Recovery Policies for Imitation Learning},
  author={Dai, Yinpei and Lee, Jayjun and Fazeli, Nima and Chai, Joyce},
  journal={arXiv preprint arXiv:2409.14674},
  year={2024}
}

Holmes-GU/RACER