Distilling Motion Planner Augmented Policies into Visual Control Policies for Robot Manipulation (CoRL 2021)

Related tags

Deep Learningreinforcement-learningroboticsmotion-planningrobot-learningrobot-manipulation
Overview

Distilling Motion Planner Augmented Policies into Visual Control Policies for Robot Manipulation

[Project website] [Paper]

This project is a PyTorch implementation of Distilling Motion Planner Augmented Policies into Visual Control Policies for Robot Manipulation, published in CoRL 2021.

Learning complex manipulation tasks in realistic, obstructed environments is a challenging problem due to hard exploration in the presence of obstacles and high-dimensional visual observations. Prior work tackles the exploration problem by integrating motion planning and reinforcement learning. However, the motion planner augmented policy requires access to state information, which is often not available in the real-world settings. To this end, we propose to distill the state-based motion planner augmented policy to a visual control policy via (1) visual behavioral cloning to remove the motion planner dependency along with its jittery motion, and (2) vision-based reinforcement learning with the guidance of the smoothed trajectories from the behavioral cloning agent. We validate our proposed approach on three manipulation tasks in obstructed environments and show its high sample-efficiency, outperforming state-of-the-art algorithms for visual policy learning.

Prerequisites

Installation

  1. Install Mujoco 2.0 and add the following environment variables into ~/.bashrc or ~/.zshrc.
# Download mujoco 2.0
$ wget https://www.roboti.us/download/mujoco200_linux.zip -O mujoco.zip
$ unzip mujoco.zip -d ~/.mujoco
$ mv ~/.mujoco/mujoco200_linux ~/.mujoco/mujoco200

# Copy mujoco license key `mjkey.txt` to `~/.mujoco`

# Add mujoco to LD_LIBRARY_PATH
$ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HOME/.mujoco/mujoco200/bin

# For GPU rendering (replace 418 with your nvidia driver version)
$ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/lib/nvidia-418

# Only for a headless server
$ export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libGLEW.so:/usr/lib/nvidia-418/libGL.so
  1. Download this repository and install python dependencies
# Install system packages
sudo apt-get install libgl1-mesa-dev libgl1-mesa-glx libosmesa6-dev patchelf libopenmpi-dev libglew-dev python3-pip python3-numpy python3-scipy

# Create out/ folder for saving/loading RL checkpoints
cd mopa-pd
mkdir out

# Create checkpoints/ folder for saving/loading BC-Visual checkpoints
mkdir checkpoints

# Install required python packages in your new env
pip install -r requirements.txt
  1. Install ompl
# Linux
sudo apt install libyaml-cpp-dev
sh ./scripts/misc/installEigen.sh #from the home directory # install Eigen
sudo apt-get install libboost-all-dev # install Boost C++ for ompl

# Mac OS
brew install libyaml yaml-cpp
brew install eigen

# Build ompl
git clone [email protected]:ompl/ompl.git ../ompl
cd ../ompl
cmake .
sudo make install

# if ompl-x.x (x.x is the version) is installed in /usr/local/include, you need to rename it to ompl
mv /usr/local/include/ompl-x.x /usr/local/include/ompl
  1. Build motion planner python wrapper
cd ./mopa-pd/motion_planners
python setup.py build_ext --inplace
  1. Configure wandb for tracking experiments (optional)
  • Sign up for a free account at https://app.wandb.ai/login?signup=true.
  • Open this file: config/__init__.py
  • Set wandb argument to True.
  • Add your username to the entity argument.
  • Add your project name to the project argument.
  1. Servers without a monitor (optional)

You may use the following code to create a virtual monitor for rendering.

# Run the next line for Ubuntu
$ sudo apt-get install xserver-xorg libglu1-mesa-dev freeglut3-dev mesa-common-dev libxmu-dev libxi-dev

# Configure nvidia-x
$ sudo nvidia-xconfig -a --use-display-device=None --virtual=1280x1024

# Launch a virtual display
$ sudo /usr/bin/X :1 &

# Run a command with DISPLAY=:1
DISPLAY=:1 <command>

Available environments

SawyerPushObstacle-v0 SawyerLiftObstacle-v0 SawyerAssemblyObstacle-v0
Sawyer Push Sawyer Lift Sawyer Assembly

How to run experiments

Launch a virtual display (only for a headless server)

sudo /usr/bin/X :1 &

MoPA-RL

# train MoPA-RL policy
sh ./scripts/3d/assembly/mopa.sh 0 1234
sh ./scripts/3d/lift/mopa.sh 0 1234
sh ./scripts/3d/push/mopa.sh 0 1234

# evaluate MoPA-RL policy
sh ./scripts/3d/assembly/mopa_eval.sh 0 1234
sh ./scripts/3d/lift/mopa_eval.sh 0 1234
sh ./scripts/3d/push/mopa_eval.sh 0 1234

# generate MoPA-RL data for BC-Visual using trained MoPA-RL's checkpoint
sh ./scripts/3d/assembly/run_multiple_sh.sh
sh ./scripts/3d/lift/run_multiple_sh.sh
sh ./scripts/3d/push/run_multiple_sh.sh

BC-Visual

# pre-process MoPA-RL data
python util/state_img_preprocessing.py

cd rl # must be inside rl folder to execute the following commands

# bc_visual_args.py is the config for training and evaluating BC-Visual

# train BC-Visual 
python behavioral_cloning_visual.py

# evaluate BC-Visual
python evaluate_bc_visual.py

Baselines and Ours

  • Sawyer Push
###### Training
sh ./scripts/3d/push/bcrl_stochastic_two_buffers.sh 0 1234 # Ours
sh ./scripts/3d/push/bcrl_stochastic_two_buffers_mopa.sh 0 1234 # Ours (w/o BC Smoothing)
sh ./scripts/3d/push/bcrl_stochastic_two_buffers_dr.sh 0 1234 # Ours (w DR)
sh ./scripts/3d/push/bcrl_mopa_sota.sh 0 1234 # CoL 
sh ./scripts/3d/push/bcrl_sota.sh 0 1234 # CoL (w BC Smoothing) 
sh ./scripts/3d/push/mopa_asym.sh 0 1234 # MoPA Asym. SAC
sh ./scripts/3d/push/bcrl_stochastic_randweights.sh 0 1234 # Asym. SAC

###### Evaluation
sh ./scripts/3d/push/bcrl_stochastic_two_buffers_eval.sh 0 1234 # Ours
sh ./scripts/3d/push/bcrl_stochastic_two_buffers_mopa_eval.sh 0 1234 # Ours (w/o BC Smoothing)
sh ./scripts/3d/push/bcrl_stochastic_two_buffers_dr_eval.sh 0 1234 # Ours (w DR)
sh ./scripts/3d/push/bcrl_mopa_sota_eval.sh 0 1234 # CoL
sh ./scripts/3d/push/bcrl_sota_eval.sh 0 1234 # CoL (w BC Smoothing)
sh ./scripts/3d/push/mopa_asym_eval.sh 0 1234  # MoPA Asym. SAC
sh ./scripts/3d/push/bcrl_stochastic_randweights_eval.sh 0 1234 # Asym. SAC
  • Sawyer Lift
###### Training
sh ./scripts/3d/lift/bcrl_stochastic_two_buffers.sh 0 1234 # Ours
sh ./scripts/3d/lift/bcrl_stochastic_two_buffers_mopa.sh 0 1234 # Ours (w/o BC Smoothing)
sh ./scripts/3d/lift/bcrl_stochastic_two_buffers_dr.sh 0 1234 # Ours (w DR)
sh ./scripts/3d/lift/bcrl_mopa_sota.sh 0 1234 # CoL
sh ./scripts/3d/lift/bcrl_sota.sh 0 1234 # CoL (w BC Smoothing) 
sh ./scripts/3d/lift/mopa_asym.sh 0 1234 # MoPA Asym. SAC
sh ./scripts/3d/lift/bcrl_stochastic_randweights.sh 0 1234 # Asym. SAC

###### Evaluation
sh ./scripts/3d/lift/bcrl_stochastic_two_buffers_eval.sh 0 1234 # Ours
sh ./scripts/3d/lift/bcrl_stochastic_two_buffers_mopa_eval.sh 0 1234 # Ours (w/o BC Smoothing)
sh ./scripts/3d/lift/bcrl_stochastic_two_buffers_dr_eval.sh 0 1234 # Ours (w DR)
sh ./scripts/3d/lift/bcrl_mopa_sota_eval.sh 0 1234 # CoL
sh ./scripts/3d/lift/bcrl_sota_eval.sh 0 1234 # CoL (w BC Smoothing)
sh ./scripts/3d/lift/mopa_asym_eval.sh 0 1234 # MoPA Asym. SAC
sh ./scripts/3d/lift/bcrl_stochastic_randweights_eval.sh 0 1234 # Asym. SAC
  • Sawyer Assembly
###### Training
sh ./scripts/3d/assembly/bcrl_stochastic_two_buffers.sh 0 1234 # Ours
sh ./scripts/3d/assembly/bcrl_stochastic_two_buffers_mopa.sh 0 1234 # Ours (w/o BC Smoothing)
sh ./scripts/3d/assembly/bcrl_stochastic_two_buffers_dr.sh 0 1234 # Ours (w DR)
sh ./scripts/3d/assembly/bcrl_mopa_sota.sh 0 1234 # CoL
sh ./scripts/3d/assembly/bcrl_sota.sh 0 1234 # CoL (w BC Smoothing)
sh ./scripts/3d/assembly/mopa_asym.sh 0 1234 # MoPA Asym. SAC
sh ./scripts/3d/assembly/bcrl_stochastic_randweights.sh 0 1234 # Asym. SAC

###### Evaluation
sh ./scripts/3d/assembly/bcrl_stochastic_two_buffers_eval.sh 0 1234 # Ours
sh ./scripts/3d/assembly/bcrl_stochastic_two_buffers_mopa_eval.sh 0 1234 # Ours (w/o BC Smoothing)
sh ./scripts/3d/assembly/bcrl_stochastic_two_buffers_dr_eval.sh 0 1234 # Ours (w DR)
sh ./scripts/3d/assembly/bcrl_mopa_sota_eval.sh 0 1234 # CoL
sh ./scripts/3d/assembly/bcrl_sota_eval.sh 0 1234 # CoL (w BC Smoothing) 
sh ./scripts/3d/assembly/mopa_asym_eval.sh 0 1234 # MoPA Asym. SAC
sh ./scripts/3d/assembly/bcrl_stochastic_randweights_eval.sh 0 1234 # Asym. SAC

Domain Randomization

To run experiments with domain randomized simulation, the following parameters can be set in config:

  • dr: set to True to train the model with domain randomization
  • dr_params_set: choose as per the training environment - ["sawyer_push, sawyer_lift, sawyer_assembly]
  • dr_eval: set to True for evaluating the domain randomization model

Directories

The structure of the repository:

  • rl: Reinforcement learning code
  • env: Environment code for simulated experiments (2D Push and all Sawyer tasks)
  • config: Configuration files
  • util: Utility code
  • motion_planners: Motion planner code from MoPA-RL
  • scripts: Scripts for all experiments

Log directories:

  • logs/rl.ENV.DATE.PREFIX.SEED:
    • cmd.sh: A command used for running a job
    • git.txt: Log gitdiff
    • prarms.json: Summary of parameters
    • video: Generated evaulation videos (every evalute_interval)
    • wandb: Training summary of W&B, like tensorboard summary
    • ckpt_*.pt: Stored checkpoints (every ckpt_interval)
    • replay_*.pt: Stored replay buffers (every ckpt_interval)

Trouble shooting

Mujoco GPU rendering

To use GPU rendering for mujoco, you need to add /usr/lib/nvidia-000 (000 should be replaced with your NVIDIA driver version) to LD_LIBRARY_PATH before installing mujoco-py. Then, during mujoco-py compilation, it will show you linuxgpuextension instead of linuxcpuextension. In Ubuntu 18.04, you may encounter an GL-related error while building mujoco-py, open venv/lib/python3.7/site-packages/mujoco_py/gl/eglshim.c and comment line 5 #include <GL/gl.h> and line 7 #include <GL/glext.h>.

Virtual display on headless machines

On servers, you don’t have a monitor. Use this to get a virtual monitor for rendering and put DISPLAY=:1 in front of a command.

# Run the next line for Ubuntu
$ sudo apt-get install xserver-xorg libglu1-mesa-dev freeglut3-dev mesa-common-dev libxmu-dev libxi-dev

# Configure nvidia-x
$ sudo nvidia-xconfig -a --use-display-device=None --virtual=1280x1024

# Launch a virtual display
$ sudo /usr/bin/X :1 &

# Run a command with DISPLAY=:1
DISPLAY=:1 <command>

pybind11-dev not found

wget http://archive.ubuntu.com/ubuntu/pool/universe/p/pybind11/pybind11-dev_2.2.4-2_all.deb
sudo apt install ./pybind11-dev_2.2.4-2_all.deb

ERROR: GLEW initalization error: Missing GL version

This issue is most likely due to running on a headless server.

Solution 1:

sudo mkdir -p /usr/lib/nvidia-000

Then add this line to ~/.bashrc file:
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/lib/nvidia-000

Solution 2:

1. First import and call mujocopy_render_hack in main.py
2. Follow the instructions in "Virtual display on headless machines" section
3. When running a script, remember to add DISPLAY:=1 <command>

/usr/bin/ld: cannot find -lGL

Source: https://stackoverflow.com/questions/33447653/usr-bin-ld-cannot-find-lgl-ubuntu-14-04

sudo rm /usr/lib/x86_64-linux-gnu/libGL.so 
sudo ln -s /usr/lib/libGL.so.1 /usr/lib/x86_64-linux-gnu/libGL.so

References

Citation

If you find this useful, please cite

@inproceedings{liu2021mopa,
  title={Distilling Motion Planner Augmented Policies into Visual Control Policies for Robot Manipulation},
  author={I-Chun Arthur Liu and Shagun Uppal and Gaurav S. Sukhatme and Joseph J. Lim and Peter Englert and Youngwoon Lee},
  booktitle={Conference on Robot Learning},
  year={2021}
}

Authors

I-Chun (Arthur) Liu*, Shagun Uppal*, Gaurav S. Sukhatme, Joseph J. Lim, Peter Englert, and Youngwoon Lee at USC CLVR and USC RESL (*Equal contribution)

Owner
Cognitive Learning for Vision and Robotics (CLVR) lab @ USC
Learning and Reasoning for Artificial Intelligence, especially focused on perception and action. Led by Professor Joseph J. Lim @ USC
Cognitive Learning for Vision and Robotics (CLVR) lab @ USC
GitHub Repository https://clvrai.com/mopa-pd
Re-implement CycleGAN in Tensorlayer

CycleGAN_Tensorlayer Re-implement CycleGAN in TensorLayer Original CycleGAN Improved CycleGAN with resize-convolution Prerequisites: TensorLayer Tenso

89 Aug 15, 2022
Learning to Reconstruct 3D Non-Cuboid Room Layout from a Single RGB Image

NonCuboidRoom Paper Learning to Reconstruct 3D Non-Cuboid Room Layout from a Single RGB Image Cheng Yang*, Jia Zheng*, Xili Dai, Rui Tang, Yi Ma, Xiao

67 Dec 15, 2022
Official PyTorch implementation of the Fishr regularization for out-of-distribution generalization

Fishr: Invariant Gradient Variances for Out-of-distribution Generalization Official PyTorch implementation of the Fishr regularization for out-of-dist

62 Dec 22, 2022
Structured Edge Detection Toolbox

################################################################### # # # Structure

Piotr Dollar 779 Jan 02, 2023
A small tool to joint picture including gif

README 做设计的时候遇到拼接长图的情况，但是发现没有什么好用的能拼接gif的工具。 于是自己写了个gif拼接小工具。 可以自动拼接gif、png和jpg等常见格式。 效果 从上至下 从下至上 从左至右 从右至左 使用 克隆仓库 git clone https://github.com/Dels

3 Dec 15, 2021
Solutions and questions for AoC2021. Merry christmas!

Advent of Code 2021 Merry christmas! 🎄 🎅 To get solutions and approximate execution times for implementations, please execute the run.py script in t

Wilhelm Ågren 5 Dec 29, 2022
Implementation of the state-of-the-art vision transformers with tensorflow

ViT Tensorflow This repository contains the tensorflow implementation of the state-of-the-art vision transformers (a category of computer vision model

Mohammadmahdi NouriBorji 2 Mar 16, 2022
This is the official source code for SLATE. We provide the code for the model, the training code, and a dataset loader for the 3D Shapes dataset. This code is implemented in Pytorch.

SLATE This is the official source code for SLATE. We provide the code for the model, the training code and a dataset loader for the 3D Shapes dataset.

Gautam Singh 66 Dec 26, 2022
Monk is a low code Deep Learning tool and a unified wrapper for Computer Vision.

Monk - A computer vision toolkit for everyone Why use Monk Issue: Want to begin learning computer vision Solution: Start with Monk's hands-on study ro

Tessellate Imaging 507 Dec 04, 2022
Pytorch implementation of AREL

Status: Archive (code is provided as-is, no updates expected) Agent-Temporal Attention for Reward Redistribution in Episodic Multi-Agent Reinforcement

8 Nov 25, 2022
Improving Query Representations for DenseRetrieval with Pseudo Relevance Feedback:A Reproducibility Study.

APR The repo for the paper Improving Query Representations for DenseRetrieval with Pseudo Relevance Feedback:A Reproducibility Study. Environment setu

ielab 8 Nov 26, 2022
PyTorch code for EMNLP 2021 paper: Don't be Contradicted with Anything! CI-ToD: Towards Benchmarking Consistency for Task-oriented Dialogue System

Don’t be Contradicted with Anything!CI-ToD: Towards Benchmarking Consistency for Task-oriented Dialogue System This repository contains the PyTorch im

Libo Qin 25 Sep 06, 2022
Bayesian optimisation library developped by Huawei Noah's Ark Library

Bayesian Optimisation Research This directory contains official implementations for Bayesian optimisation works developped by Huawei R&D, Noah's Ark L

HUAWEI Noah's Ark Lab 395 Dec 30, 2022
A Demo server serving Bert through ONNX with GPU written in Rust with <3

Demo BERT ONNX server written in rust This demo showcase the use of onnxruntime-rs on BERT with a GPU on CUDA 11 served by actix-web and tokenized wit

Xavier Tao 28 Jan 01, 2023
Implementation of PersonaGPT Dialog Model

PersonaGPT An open-domain conversational agent with many personalities PersonaGPT is an open-domain conversational agent cpable of decoding personaliz

ILLIDAN Lab 42 Jan 01, 2023
A distributed, plug-n-play algorithm for multi-robot applications with a priori non-computable objective functions

A distributed, plug-n-play algorithm for multi-robot applications with a priori non-computable objective functions Kapoutsis, A.C., Chatzichristofis,

Athanasios Ch. Kapoutsis 5 Oct 15, 2022
FedJAX is a library for developing custom Federated Learning (FL) algorithms in JAX.

FedJAX: Federated learning with JAX What is FedJAX? FedJAX is a library for developing custom Federated Learning (FL) algorithms in JAX. FedJAX priori

Google 208 Dec 14, 2022
Code and data of the EMNLP 2021 paper "Mind the Style of Text! Adversarial and Backdoor Attacks Based on Text Style Transfer"

StyleAttack Code and data of the EMNLP 2021 paper "Mind the Style of Text! Adversarial and Backdoor Attacks Based on Text Style Transfer" Prepare Pois

THUNLP 19 Nov 20, 2022
Code related to the manuscript "Averting A Crisis In Simulation-Based Inference"

Abstract We present extensive empirical evidence showing that current Bayesian simulation-based inference algorithms are inadequate for the falsificat

Montefiore Artificial Intelligence Research 3 Nov 14, 2022
A Python library for differentiable optimal control on accelerators.

A Python library for differentiable optimal control on accelerators.

Google 80 Dec 21, 2022