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

Ubuntu 18.04
Python 3.7
MuJoCo 2.0

Installation

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

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

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

Build motion planner python wrapper

cd ./mopa-pd/motion_planners
python setup.py build_ext --inplace

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.

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

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

Before running the following training scripts, please download checkpoints from (https://www.dropbox.com/sh/g8c5gz0k63qx1hl/AAD8sh_lw-JrQPAm-UoHVBUfa?dl=0) and place them inside the out/ folder.

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

This codebase is based on MoPA-RL: https://github.com/clvrai/mopa-rl
PyTorch implementation of SAC: https://github.com/clvrai/coordination
Sawyer MuJoCo environment: https://github.com/StanfordVL/robosuite
Sawyer assets: https://github.com/vikashplus/sawyer_sim

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)

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

Related tags

Overview

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

Prerequisites

Installation

Available environments

How to run experiments

Domain Randomization

Directories

Trouble shooting

Mujoco GPU rendering

Virtual display on headless machines

pybind11-dev not found

ERROR: GLEW initalization error: Missing GL version

/usr/bin/ld: cannot find -lGL

References

Citation

Authors

Owner

Cognitive Learning for Vision and Robotics (CLVR) lab @ USC

This is a clean and robust Pytorch implementation of DQN and Double DQN.

LONG-TERM SERIES FORECASTING WITH QUERYSELECTOR – EFFICIENT MODEL OF SPARSEATTENTION

Auxiliary Raw Net (ARawNet) is a ASVSpoof detection model taking both raw waveform and handcrafted features as inputs, to balance the trade-off between performance and model complexity.

object recognition with machine learning on Respberry pi

Implementation of Online Label Smoothing in PyTorch

Towards Multi-Camera 3D Human Pose Estimation in Wild Environment

A PyTorch implementation of the baseline method in Panoptic Narrative Grounding (ICCV 2021 Oral)

Strongly local p-norm-cut algorithms for semi-supervised learning and local graph clustering

deep-prae

Python scripts form performing stereo depth estimation using the HITNET model in ONNX.

some academic posters as references. May we have in-person poster session soon!

[ICCV 2021] HRegNet: A Hierarchical Network for Large-scale Outdoor LiDAR Point Cloud Registration

The official implementation of the Interspeech 2021 paper WSRGlow: A Glow-based Waveform Generative Model for Audio Super-Resolution.

Online Multi-Granularity Distillation for GAN Compression (ICCV2021)

Imaginaire - NVIDIA's Deep Imagination Team's PyTorch Library

A very simple tool to rewrite parameters such as attributes and constants for OPs in ONNX models. Simple Attribute and Constant Modifier for ONNX.

This repo will contain code to reproduce and build upon understanding transfer learning

HugsVision is a easy to use huggingface wrapper for state-of-the-art computer vision

A multi-entity Transformer for multi-agent spatiotemporal modeling.

Deepparse is a state-of-the-art library for parsing multinational street addresses using deep learning