RoboDesk A Multi-Task Reinforcement Learning Benchmark

Related tags

Deep Learningrobodesk
Overview

RoboDesk

PyPI

A Multi-Task Reinforcement Learning Benchmark

Robodesk Banner

If you find this open source release useful, please reference in your paper:

@misc{kannan2021robodesk,
  author = {Harini Kannan and Danijar Hafner and Chelsea Finn and Dumitru Erhan},
  title = {RoboDesk: A Multi-Task Reinforcement Learning Benchmark},
  year = {2021},
  howpublished = {\url{https://github.com/google-research/robodesk}},
}

Highlights

  • Diversity: RoboDesk includes 9 diverse tasks that test for a variety of different behaviors within the same environment, making it useful for evaluating transfer, multi-task learning, and global exploration.
  • Complexity: The high-dimensional image inputs contain objects of different shapes and colors, whose initial positions are randomized to avoid naive memorization and require learning algorithms to generalize.
  • Robustness: We carefully designed and tested RoboDesk to ensure fast and stable physics simulation. This avoids objects from intersecting, getting stuck, or quickly flying away, a common problem with some existing environments.
  • Lightweight: RoboDesk comes as a self-contained Python package with few dependencies. The source code is clean and pragmatic, making it a useful blueprint for creating new MuJoCo environments.

Training Agents

Installation: pip3 install -U robodesk

The environment follows the OpenAI Gym interface:

import robodesk

env = robodesk.RoboDesk(seed=0)
obs = env.reset()
assert obs.shape == (64, 64, 3)

done = False
while not done:
  action = env.action_space.sample()
  obs, reward, done, info = env.step(action)

Tasks

Robodesk Tasks

The behaviors above were learned using the Dreamer agent. These policies have been learned from scratch and only from pixels, not proprioceptive states.

Task Description
open_slide Push the sliding door all the way to the right, navigating around the other objects.
open_drawer Pull the dark brown drawer all the way open.
push_green Push the green button to turn the green light on.
stack_blocks Stack the upright blue block on top of the flat green block.
upright_block_off_table Push the blue upright block off the table.
flat_block_in_bin Push the green flat block into the blue bin.
flat_block_in_shelf Push the green flat block into the shelf, navigating around the other blocks.
lift_upright_block Grasp the blue upright block and lift it above the table.
lift_ball Grasp the magenta ball and lift it above the table.

Environment Details

Constructor

robodesk.RoboDesk(task='open_slide', reward='dense', action_repeat=1, episode_length=500, image_size=64)
Parameter Description
task Available tasks are open_slide, open_drawer, push_green, stack, upright_block_off_table, flat_block_in_bin, flat_block_in_shelf, lift_upright_block, lift_ball.
reward Available reward types are dense, sparse, success. Success gives only the first sparse reward during the episode, useful for computing success rates during evaluation.
action_repeat Reduces the control frequency by applying each action multiple times. This is faster than using an environment wrapper because only the needed images are rendered.
episode_length Time limit for the episode, can be None.
image_size Size of the image observations in pixels, used for both height and width.

Reward

All rewards are bound between 0 and 1. There are three types of rewards available:

  • Dense rewards are based on Euclidean distances between the objects and their target positions and can include additional terms, for example to encourage the arm to reach the object. These are the easiest rewards for learning.
  • Sparse rewards are either 0 or 1 based on whether the target object is in the target area or not, according to a fixed threshold. Learning from sparse rewards is more challenging.
  • Success rewards are equivalent to the sparse rewards, except that only the first reward is given during each episode. As a result, an episode return of 0 means failure and 1 means sucess at the task. This should only be used during evaluation.

Termination

Episodes end after 500 time steps by default. There are no early terminations.

Observation Space

Each observation is a dictionary that contains the current image, as well as additional information. For the standard benchmark, only the image should be used for learning. The observation dictionary contains the following keys:

Key Space
image Box(0, 255, (64, 64, 3), np.uint8)
qpos_robot Box(-np.inf, np.inf, (9,), np.float32)
qvel_robot Box(-np.inf, np.inf, (9,), np.float32)
qpos_objects Box(-np.inf, np.inf, (26,), np.float32)
qvel_objects Box(-np.inf, np.inf, (26,), np.float32)
end_effector Box(-np.inf, np.inf, (3,), np.float32)

Action Space

RoboDesk uses end effector control with a simple bounded action space:

Box(-1, 1, (5,), np.float32)

Acknowledgements

We thank Ben Eysenbach and Debidatta Dwibedi for their helpful feedback.

Our benchmark builds upon previously open-sourced work. We build upon the desk XMLs first introduced in [1], the Franka XMLs open-sourced in [2], and the Franka meshes open-sourced in [3].

Questions

Please open an issue on Github.

Disclaimer: This is not an official Google product.

Owner
Google Research
Google Research
QQ Browser 2021 AI Algorithm Competition Track 1 1st Place Program

QQ Browser 2021 AI Algorithm Competition Track 1 1st Place Program

249 Jan 03, 2023
Code for 'Single Image 3D Shape Retrieval via Cross-Modal Instance and Category Contrastive Learning', ICCV 2021

CMIC-Retrieval Code for Single Image 3D Shape Retrieval via Cross-Modal Instance and Category Contrastive Learning. ICCV 2021. Introduction In this wo

42 Nov 17, 2022
Code for "Hierarchical Skills for Efficient Exploration" HSD-3 Algorithm and Baselines

Hierarchical Skills for Efficient Exploration This is the source code release for the paper Hierarchical Skills for Efficient Exploration. It contains

Facebook Research 38 Dec 06, 2022
Link prediction using Multiple Order Local Information (MOLI)

Understanding the network formation pattern for better link prediction Authors: [e

Wu Lab 0 Oct 18, 2021
A Transformer-Based Siamese Network for Change Detection

ChangeFormer: A Transformer-Based Siamese Network for Change Detection (Under review at IGARSS-2022) Wele Gedara Chaminda Bandara, Vishal M. Patel Her

Wele Gedara Chaminda Bandara 214 Dec 29, 2022
RLBot Python bindings for the Rust crate rl_ball_sym

RLBot Python bindings for rl_ball_sym 0.6 Prerequisites: Rust & Cargo Build Tools for Visual Studio RLBot - Verify that the file %localappdata%\RLBotG

Eric Veilleux 2 Nov 25, 2022
Tensorflow/Keras Plug-N-Play Deep Learning Models Compilation

DeepBay This project was created with the objective of compile Machine Learning Architectures created using Tensorflow or Keras. The architectures mus

Whitman Bohorquez 4 Sep 26, 2022
structured-generative-modeling

This repository contains the implementation for the paper Information Theoretic StructuredGenerative Modeling, Specially thanks for the open-source co

0 Oct 11, 2021
the code for paper "Energy-Based Open-World Uncertainty Modeling for Confidence Calibration"

EOW-Softmax This code is for the paper "Energy-Based Open-World Uncertainty Modeling for Confidence Calibration". Accepted by ICCV21. Usage Commnd exa

Yezhen Wang 36 Dec 02, 2022
Bu repo SAHI uygulamasını mantığını öğreniyoruz.

SAHI-Learn: SAHI'den Beraber Kodlamak İster Misiniz Herkese merhabalar ben Kadir Nar. SAHI kütüphanesine gönüllü geliştiriciyim. Bu repo SAHI kütüphan

Kadir Nar 11 Aug 22, 2022
Official PyTorch implementation of N-ImageNet: Towards Robust, Fine-Grained Object Recognition with Event Cameras (ICCV 2021)

N-ImageNet: Towards Robust, Fine-Grained Object Recognition with Event Cameras Official PyTorch implementation of N-ImageNet: Towards Robust, Fine-Gra

32 Dec 26, 2022
Defending graph neural networks against adversarial attacks (NeurIPS 2020)

GNNGuard: Defending Graph Neural Networks against Adversarial Attacks Authors: Xiang Zhang ( Zitnik Lab @ Harvard 44 Dec 07, 2022

This repository is based on Ultralytics/yolov5, with adjustments to enable polygon prediction boxes.

Polygon-Yolov5 This repository is based on Ultralytics/yolov5, with adjustments to enable polygon prediction boxes. Section I. Description The codes a

xinzelee 226 Jan 05, 2023
Training RNNs as Fast as CNNs

News SRU++, a new SRU variant, is released. [tech report] [blog] The experimental code and SRU++ implementation are available on the dev branch which

ASAPP Research 2.1k Jan 01, 2023
I will implement Fastai in each projects present in this repository.

DEEP LEARNING FOR CODERS WITH FASTAI AND PYTORCH The repository contains a list of the projects which I have worked on while reading the book Deep Lea

Thinam Tamang 43 Dec 20, 2022
A Sign Language detection project using Mediapipe landmark detection and Tensorflow LSTM's

sign-language-detection A Sign Language detection project using Mediapipe landmark detection and Tensorflow LSTM. The project is built for a vocabular

Hashim 4 Feb 06, 2022
Code for our paper "Interactive Analysis of CNN Robustness"

Perturber Code for our paper "Interactive Analysis of CNN Robustness" Datasets Feature visualizations: Google Drive Fine-tuning checkpoints as saved m

Stefan Sietzen 0 Aug 17, 2021
[PAMI 2020] Show, Match and Segment: Joint Weakly Supervised Learning of Semantic Matching and Object Co-segmentation

Show, Match and Segment: Joint Weakly Supervised Learning of Semantic Matching and Object Co-segmentation This repository contains the source code for

Yun-Chun Chen 60 Nov 25, 2022
A model that attempts to learn and benefit from data collected on card counting.

A model that attempts to learn and benefit from data collected on card counting. A decision tree like model is built to win more often than loose and increase the bet of the player appropriately to c

1 Dec 17, 2021
mmfewshot is an open source few shot learning toolbox based on PyTorch

OpenMMLab FewShot Learning Toolbox and Benchmark

OpenMMLab 514 Dec 28, 2022