SGRv2 & SGR

This is the official repository for SGRv2 and SGR.

SGRv2

Leveraging Locality to Boost Sample Efficiency in Robotic Manipulation
Tong Zhang, Yingdong Hu, Jiacheng You, Yang Gao

CoRL, 2024

Left: Sample efficiency of SGRv2. Top Right: Overview of simulation results. Bottom Right: Tasks of the 3 simulation benchmarks.

SGR

A Universal Semantic-Geometric Representation for Robotic Manipulation
Tong Zhang*, Yingdong Hu*, Hanchen Cui, Hang Zhao, Yang Gao

CoRL, 2023

Leveraging semantic information from massive 2D images and geometric information from 3D point clouds, we present Semantic-Geometric Representation (SGR) that enables the robots to solve a range of simulated and real-world manipulation tasks.

Keyframe Control on RLBench

Install

• Tested (Recommended) Versions: Python 3.8. We used CUDA 11.1.

• Step 1: We recommend using conda and creating a virtual environment.

conda create --name sgr python=3.8
conda activate sgr

• Step 2: Install PyTorch. Make sure the PyTorch version is compatible with the CUDA version. One recommended version compatible with CUDA 11.1 can be installed with the following command. More instructions to install PyTorch can be found here.

conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch

• Step 3: Install CoppeliaSim. PyRep requires version 4.1 of CoppeliaSim. Download and unzip CoppeliaSim:
• Ubuntu 16.04
• Ubuntu 18.04
• Ubuntu 20.04

Once you have downloaded CoppeliaSim, add the following to your ~/.bashrc file. (NOTE: the 'EDIT ME' in the first line)

export COPPELIASIM_ROOT=<EDIT ME>/PATH/TO/COPPELIASIM/INSTALL/DIR
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$COPPELIASIM_ROOT
export QT_QPA_PLATFORM_PLUGIN_PATH=$COPPELIASIM_ROOT

Remember to source your .bashrc (source ~/.bashrc) or .zshrc (source ~/.zshrc) after this.

• Step 4: Install PyRep. Once you have install CoppeliaSim, you can pull PyRep from git:

cd <install_dir>
git clone https://github.com/stepjam/PyRep.git
cd PyRep

Then install the python library:

pip install -r requirements.txt
pip install .

• Step 5: Clone the SGR repository with the submodules using the following command.

cd <install_dir>
git clone --recurse-submodules git@github.com:TongZhangTHU/sgr.git 
cd sgr 
git submodule update --init

Now, locally install my RLBench fork, my YARR fork, my openpoints fork, and other libraries using the following command. Make sure you are in folder sgr.

pip install -r libs/RLBench/requirements.txt
pip install -e libs/RLBench 
pip install -r libs/YARR/requirements.txt
pip install -e libs/YARR 
pip install git+https://github.com/openai/CLIP.git
pip install -r requirements.txt

For running RLBench/CoppeliaSim in headless mode, please refer to the RLBench official repo and CoppeliaSim forums.

• Step 6: Install the C++ extensions, the PointNet++ library. These are used to speed up the farthest point sampling (FPS).

cd openpoints/cpp/pointnet2_batch
pip install .
cd ../
cd chamfer_dist
pip install . --user
cd ../../../

Data Generation

We utilize the tools in libs/RLBench/tools/dataset_generator.py to generate data. In SGRv2, since each task utilizes only 5 demonstrations, we use the first variation for tasks with multiple variations, in contrast to PerAct, which combines all variations.

Below is an example of how to generate data. For more details, see scripts/gen_data.sh. These commands can be executed in parallel for multiple tasks.

python libs/RLBench/tools/dataset_generator.py \
                          --save_path=data/train \
                          --tasks=open_microwave \
                          --image_size=128,128 \
                          --renderer=opengl \
                          --episodes_per_task=100 \
                          --variations=1 \
                          --all_variations=False

Training and Evaluatation

The following is a guide for training everything from scratch. All tasks follow a 4-phase workflow:

1. Generate train and test datasets using libs/RLBench/tools/dataset_generator.py.
2. Train SGRv2 or SGR using train.py with 5 demonstrations per task for 20,000 iterations, saving checkpoints every 800 iterations. If training with more demonstrations, it is recommended to increase the number of iterations accordingly.
3. Run evaluation using eval.py with framework.eval_type=missing5 and framework.eval_episodes=50 to assess the last 5 checkpoints across 50 episodes on test data, and save the results in eval_data.csv.
4. Repeat steps 2 and 3 with 3 seeds and report the average results.

Below is an example of how to train and evaluate SGRv2 and SGR. For more details, see scripts/run_sgrv2.sh and scripts/run_sgrv1.sh. It is recommended to run multiple seeds to reduce variance. And it' better to increase training iterations if number of demos increase.

Train SGRv2 on open_microwave with 5 demos for 20000 iterations:

CUDA_VISIBLE_DEVICES=0 python train.py  rlbench.tasks=open_microwave \
                                        rlbench.demos=5 \
                                        rlbench.demo_path=data/train \
                                        replay.batch_size=16 \
                                        framework.start_seed=0 \
                                        framework.save_freq=800 \
                                        framework.training_iterations=20000 \
                                        method=SGR \
                                        method.color_drop=0.4 \
                                        method.tag=sgrv2-demos_5-iter_20000 \
                                        model=pointnext-xl_seg \
                                        model.cls_args.mlps=[256] \
                                        model.cls_args.num_classes=256

Train SGRv1 on open_microwave with 5 demos for 20000 iterations:

CUDA_VISIBLE_DEVICES=0 python train.py  rlbench.tasks=open_microwave \
                                        rlbench.demos=5 \
                                        rlbench.demo_path=data/train \
                                        replay.batch_size=32 \
                                        framework.start_seed=0 \
                                        framework.save_freq=800 \
                                        framework.training_iterations=20000 \
                                        method=SGR \
                                        method.color_drop=0.2 \
                                        method.tag=sgrv1-demos_5-iter_20000 \
                                        model=pointnext-s_cls

Evalute last 5 checkpoints of SGRv2 on open_microwave for 50 episodes:

CUDA_VISIBLE_DEVICES=0 python eval.py rlbench.tasks=open_microwave \
                                      rlbench.demo_path=data/test \
                                      framework.start_seed=0 \
                                      framework.eval_type=missing5 \
                                      framework.eval_envs=5 \
                                      framework.eval_episodes=50 \
                                      method.name=SGR \
                                      method.tag=sgrv2-demos_5-iter_20000 \
                                      model.name=pointnext-xl_seg

Evalute last 5 checkpoints of SGRv1 on open_microwave for 50 episodes:

CUDA_VISIBLE_DEVICES=0 python eval.py rlbench.tasks=open_microwave \
                                      rlbench.demo_path=data/test \
                                      framework.start_seed=0 \
                                      framework.eval_type=missing5 \
                                      framework.eval_envs=5 \
                                      framework.eval_episodes=50 \
                                      method.name=SGR \
                                      method.tag=sgrv1-demos_5-iter_20000 \
                                      model.name=pointnext-s_cls

TODO

• [ ] Scripts for baseline models, including PerAct, PointNeXt, and R3M.
• [ ] Scripts for ablation studies.
• [ ] The implementation of dense control.

Acknowledgement

We sincerely thank the authors of the following repositories for sharing their code. - PerAct - PyRep - RLBench - YARR - RVT

Citation

If you find our work useful, please consider citing:

@article{zhang2024leveraging,
    title={Leveraging Locality to Boost Sample Efficiency in Robotic Manipulation},
    author={Zhang, Tong and Hu, Yingdong and You, Jiacheng and Gao, Yang},
    journal={arXiv preprint arXiv:2406.10615},
    year={2024}
  }

@article{zhang2023universal,
    title={A Universal Semantic-Geometric Representation for Robotic Manipulation},
    author={Zhang, Tong and Hu, Yingdong and Cui, Hanchen and Zhao, Hang and Gao, Yang},
    journal={arXiv preprint arXiv:2306.10474},
    year={2023}
  }