[NeurIPS 2021] PyTorch Code for Accelerating Robotic Reinforcement Learning with Parameterized Action Primitives

Last update: Dec 27, 2022

Overview

Robot Action Primitives (RAPS)

This repository is the official implementation of Accelerating Robotic Reinforcement Learning via Parameterized Action Primitives (RAPS).

[Project Website]

Murtaza Dalal, Deepak Pathak*, Ruslan Salakhutdinov*
(* equal advising)

CMU

If you find this work useful in your research, please cite:

@inproceedings{dalal2021raps,
    Author = {Dalal, Murtaza and Pathak, Deepak and
              Salakhutdinov, Ruslan},
    Title = {Accelerating Robotic Reinforcement Learning via Parameterized Action Primitives},
    Booktitle = {NeurIPS},
    Year = {2021}
}

Requirements

To install dependencies, please run the following commands:

sudo apt-get update
sudo apt-get install curl \
    git \
    libgl1-mesa-dev \
    libgl1-mesa-glx \
    libglew-dev \
    libosmesa6-dev \
    software-properties-common \
    net-tools \
    unzip \
    vim \
    virtualenv \
    wget \
    xpra \
    xserver-xorg-dev
sudo apt-get install libglfw3-dev libgles2-mesa-dev patchelf
sudo mkdir /usr/lib/nvidia-000

Please add the following to your bashrc:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:~/.mujoco/mujoco200/bin
export MUJOCO_GL='egl'
export MKL_THREADING_LAYER=GNU
export D4RL_SUPPRESS_IMPORT_ERROR='1'
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/lib/nvidia-000

To install python requirements:

conda create -n raps python=3.7
conda activate raps
./setup_python_env.sh <absolute path to raps>

Training and Evaluation

Kitchen

Prior to running any experiments, make sure to run cd /path/to/raps/rlkit

single task env names:

microwave
kettle
slide_cabinet
hinge_cabinet
light_switch
top_left_burner

multi task env names:

microwave_kettle_light_top_left_burner //Sequential Multi Task 1
hinge_slide_bottom_left_burner_light //Sequential Multi Task 2

To train RAPS with Dreamer on any single task kitchen environment, run:

python experiments/kitchen/dreamer/dreamer_v2_single_task_primitives.py --mode here_no_doodad --exp_prefix <> --env <env name>

To train RAPS with Dreamer on the multi task kitchen environments, run:

python experiments/kitchen/dreamer/dreamer_v2_multi_task_primitives.py --mode here_no_doodad --exp_prefix <> --env <env name>

To train Raw Actions with Dreamer on any kitchen environment

python experiments/kitchen/dreamer/dreamer_v2_raw_actions.py --mode here_no_doodad --exp_prefix <> --env <env name>

To train RAPS with RAD on any single task kitchen environment

python experiments/kitchen/rad/rad_single_task_primitives.py --mode here_no_doodad --exp_prefix <> --env <env name>

To train RAPS with RAD on any multi task kitchen environment

python experiments/kitchen/rad/rad_multi_task_primitives.py --mode here_no_doodad --exp_prefix <> --env <env name>

To train Raw Actions with RAD on any kitchen environment

python experiments/kitchen/rad/rad_raw_actions.py --mode here_no_doodad --exp_prefix <> --env <env name>

To train RAPS with PPO on any single task kitchen environment

python experiments/kitchen/ppo/ppo_single_task_primitives.py --mode here_no_doodad --exp_prefix <> --env <env name>

To train RAPS with PPO on any multi task kitchen environment

python experiments/kitchen/ppo/ppo_multi_task_primitives.py --mode here_no_doodad --exp_prefix <> --env <env name>

To train Raw Actions with PPO on any kitchen environment

python experiments/kitchen/ppo/ppo_raw_actions.py --mode here_no_doodad --exp_prefix <> --env <env name>

Metaworld

single task env names

drawer-close-v2
soccer-v2
peg-unplug-side-v2
sweep-into-v2
assembly-v2
disassemble-v2

To train RAPS with Dreamer on any metaworld environment

python experiments/metaworld/dreamer/dreamer_v2_single_task_primitives.py --mode here_no_doodad --exp_prefix <> --env <env name>

To train Raw Actions with Dreamer on any metaworld environment

python experiments/metaworld/dreamer/dreamer_v2_single_task_raw_actions.py --mode here_no_doodad --exp_prefix <> --env <env name>

Robosuite

To train RAPS with Dreamer on an Robosuite Lift

python experiments/robosuite/dreamer/dreamer_v2_single_task_primitives_lift.py --mode here_no_doodad --exp_prefix <>

To train Raw Actions with Dreamer on an Robosuite Lift

python experiments/robosuite/dreamer/dreamer_v2_single_task_raw_actions_lift.py --mode here_no_doodad --exp_prefix <>

To train RAPS with Dreamer on an Robosuite Door

python experiments/robosuite/dreamer/dreamer_v2_single_task_primitives_door.py --mode here_no_doodad --exp_prefix <>

To train Raw Actions with Dreamer on an Robosuite Door

python experiments/robosuite/dreamer/dreamer_v2_single_task_raw_actions_door.py --mode here_no_doodad --exp_prefix <>

Learning Curve visualization

cd /path/to/raps/rlkit
python ../viskit/viskit/frontend.py data/<exp_prefix> //open localhost:5000 to view

[NeurIPS 2021] PyTorch Code for Accelerating Robotic Reinforcement Learning with Parameterized Action Primitives

Related tags

Overview

Robot Action Primitives (RAPS)

[Project Website]

Requirements

Training and Evaluation

Kitchen

Metaworld

Robosuite

Learning Curve visualization

Owner

Murtaza Dalal

Tutorials, assignments, and competitions for MIT Deep Learning related courses.

Supervision Exists Everywhere: A Data Efficient Contrastive Language-Image Pre-training Paradigm

Unofficial PyTorch implementation of MobileViT.

My take on a practical implementation of Linformer for Pytorch.

Global Filter Networks for Image Classification

Code for the paper "Attention Approximates Sparse Distributed Memory"

Cluster-GCN: An Efficient Algorithm for Training Deep and Large Graph Convolutional Networks

CS_Final_Metal_surface_detection - This is a final project for CoderSchool Machine Learning bootcamp on 29/12/2021.

implementation of paper - You Only Learn One Representation: Unified Network for Multiple Tasks

Source code of SIGIR2021 Paper 'One Chatbot Per Person: Creating Personalized Chatbots based on Implicit Profiles'

Torchserve server using a YoloV5 model running on docker with GPU and static batch inference to perform production ready inference.

Bag of Tricks for Natural Policy Gradient Reinforcement Learning

Calibrated Hyperspectral Image Reconstruction via Graph-based Self-Tuning Network.

Pythonic particle-based (super-droplet) warm-rain/aqueous-chemistry cloud microphysics package with box, parcel & 1D/2D prescribed-flow examples in Python, Julia and Matlab

Forest R-CNN: Large-Vocabulary Long-Tailed Object Detection and Instance Segmentation (ACM MM 2020)

Implementation of the bachelor's thesis "Real-time stock predictions with deep learning and news scraping".

Official code release for 3DV 2021 paper Human Performance Capture from Monocular Video in the Wild.

Trajectory Extraction of road users via Traffic Camera

Self-Supervised Pillar Motion Learning for Autonomous Driving (CVPR 2021)

A framework for joint super-resolution and image synthesis, without requiring real training data