RGB-stacking ๐Ÿ›‘ ๐ŸŸฉ ๐Ÿ”ท for robotic manipulation

Overview

RGB-stacking ๐Ÿ›‘ ๐ŸŸฉ ๐Ÿ”ท for robotic manipulation

BLOG | PAPER | VIDEO

Beyond Pick-and-Place: Tackling Robotic Stacking of Diverse Shapes,
Alex X. Lee*, Coline Devin*, Yuxiang Zhou*, Thomas Lampe*, Konstantinos Bousmalis*, Jost Tobias Springenberg*, Arunkumar Byravan, Abbas Abdolmaleki, Nimrod Gileadi, David Khosid, Claudio Fantacci, Jose Enrique Chen, Akhil Raju, Rae Jeong, Michael Neunert, Antoine Laurens, Stefano Saliceti, Federico Casarini, Martin Riedmiller, Raia Hadsell, Francesco Nori.
In Conference on Robot Learning (CoRL), 2021.

The RGB environment

This repository contains an implementation of the simulation environment described in the paper "Beyond Pick-and-Place: Tackling robotic stacking of diverse shapes". Note that this is a re-implementation of the environment (to remove dependencies on internal libraries). As a result, not all the features described in the paper are available at this point. Noticeably, domain randomization is not included in this release. We also aim to provide reference performance metrics of trained policies on this environment in the near future.

In this environment, the agent controls a robot arm with a parallel gripper above a basket, which contains three objects โ€” one red, one green, and one blue, hence the name RGB. The agent's task is to stack the red object on top of the blue object, within 20 seconds, while the green object serves as an obstacle and distraction. The agent controls the robot using a 4D Cartesian controller. The controlled DOFs are x, y, z and rotation around the z axis. The simulation is a MuJoCo environment built using the Modular Manipulation (MoMa) framework.

Corresponding method

The RGB-stacking paper "Beyond Pick-and-Place: Tackling robotic stacking of diverse shapes" also contains a description and thorough evaluation of our initial solution to both the 'Skill Mastery' (training on the 5 designated test triplets and evaluating on them) and the 'Skill Generalization' (training on triplets of training objects and evaluating on the 5 test triplets). Our approach was to first train a state-based policy in simulation via a standard RL algorithm (we used MPO) followed by interactive distillation of the state-based policy into a vision-based policy (using a domain randomized version of the environment) that we then deployed to the robot via zero-shot sim-to-real transfer. We finally improved the policy further via offline RL based on data collected from the sim-to-real policy (we used CRR). For details on our method and the results please consult the paper.

Installing and visualizing the environment

Please ensure that you have a working MuJoCo200 installation and a valid MuJoCo licence.

  1. Clone this repository:

    git clone https://github.com/deepmind/rgb_stacking.git
    cd rgb_stacking
  2. Prepare a Python 3 environment - venv is recommended.

    python3 -m venv rgb_stacking_venv
    source rgb_stacking_venv/bin/activate
  3. Install dependencies:

    pip install -r requirements.txt
  4. Run the environment viewer:

    python -m rgb_stacking.main

Step 2-4 can also be done by running the run.sh script:

./run.sh

Specifying the object triplet

The default environment will load with Triplet 4 (see Sect. 3.2.1 in the paper). If you wish to use a different triplet you can use the following commands:

from rgb_stacking import environment

env = environment.rgb_stacking(object_triplet=NAME_OF_SET)

The possible NAME_OF_SET are:

  • rgb_test_triplet{i} where i is one of 1, 2, 3, 4, 5: Loads test triplet i.
  • rgb_test_random: Randomly loads one of the 5 test triplets.
  • rgb_train_random: Triplet comprised of blocks from the training set.
  • rgb_heldout_random: Triplet comprised of blocks from the held-out set.

For more information on the blocks and the possible options, please refer to the rgb_objects repository.

Specifying the observation space

By default, the observations exposed by the environment are only the ones we used for training our state-based agents. To use another set of observations please use the following code snippet:

from rgb_stacking import environment

env = environment.rgb_stacking(
    observations=environment.ObservationSet.CHOSEN_SET)

The possible CHOSEN_SET are:

  • STATE_ONLY: Only the state observations, used for training expert policies from state in simulation (stage 1).
  • VISION_ONLY: Only image observations.
  • ALL: All observations.
  • INTERACTIVE_IMITATION_LEARNING: Pair of image observations and a subset of proprioception observations, used for interactive imitation learning (stage 2).
  • OFFLINE_POLICY_IMPROVEMENT: Pair of image observations and a subset of proprioception observations, used for the one-step offline policy improvement (stage 3).

Real RGB-Stacking Environment: CAD models and assembly instructions

The CAD model of the setup is available in onshape.

We also provide the following documents for the assembly of the real cell:

  • Assembly instructions for the basket.
  • Assembly instructions for the robot.
  • Assembly instructions for the cell.
  • The bill of materials of all the necessary parts.
  • A diagram with the wiring of cell.

The RGB-objects themselves can be 3D-printed using the STLs available in the rgb_objects repository.

Citing

If you use rgb_stacking in your work, please cite the accompanying paper:

@inproceedings{lee2021rgbstacking,
    title={Beyond Pick-and-Place: Tackling Robotic Stacking of Diverse Shapes},
    author={Alex X. Lee and
            Coline Devin and
            Yuxiang Zhou and
            Thomas Lampe and
            Konstantinos Bousmalis and
            Jost Tobias Springenberg and
            Arunkumar Byravan and
            Abbas Abdolmaleki and
            Nimrod Gileadi and
            David Khosid and
            Claudio Fantacci and
            Jose Enrique Chen and
            Akhil Raju and
            Rae Jeong and
            Michael Neunert and
            Antoine Laurens and
            Stefano Saliceti and
            Federico Casarini and
            Martin Riedmiller and
            Raia Hadsell and
            Francesco Nori},
    booktitle={Conference on Robot Learning (CoRL)},
    year={2021},
    url={https://openreview.net/forum?id=U0Q8CrtBJxJ}
}
Owner
DeepMind
DeepMind
Code of the paper "Deep Human Dynamics Prior" in ACM MM 2021.

Code of the paper "Deep Human Dynamics Prior" in ACM MM 2021. Figure 1: In the process of motion capture (mocap), some joints or even the whole human

Shinny cui 3 Oct 31, 2022
Codes for "CSDI: Conditional Score-based Diffusion Models for Probabilistic Time Series Imputation"

CSDI This is the github repository for the NeurIPS 2021 paper "CSDI: Conditional Score-based Diffusion Models for Probabilistic Time Series Imputation

106 Jan 04, 2023
AI-UPV at IberLEF-2021 EXIST task: Sexism Prediction in Spanish and English Tweets Using Monolingual and Multilingual BERT and Ensemble Models

AI-UPV at IberLEF-2021 EXIST task: Sexism Prediction in Spanish and English Tweets Using Monolingual and Multilingual BERT and Ensemble Models Descrip

Angel de Paula 1 Jun 08, 2022
Implementation of "Unsupervised Domain Adaptive 3D Detection with Multi-Level Consistency"

Unsupervised Domain Adaptive 3D Detection with Multi-Level Consistency (ICCV2021) Paper Link: https://arxiv.org/abs/2107.11355 This implementation bui

32 Nov 17, 2022
Implementation of popular bandit algorithms in batch environments.

batch-bandits Implementation of popular bandit algorithms in batch environments. Source code to our paper "The Impact of Batch Learning in Stochastic

Danil Provodin 2 Sep 11, 2022
Official repo of the paper "Surface Form Competition: Why the Highest Probability Answer Isn't Always Right"

Surface Form Competition This is the official repo of the paper "Surface Form Competition: Why the Highest Probability Answer Isn't Always Right" We p

Peter West 46 Dec 23, 2022
KoCLIP: Korean port of OpenAI CLIP, in Flax

KoCLIP This repository contains code for KoCLIP, a Korean port of OpenAI's CLIP. This project was conducted as part of Hugging Face's Flax/JAX communi

Jake Tae 100 Jan 02, 2023
Codes accompanying the paper "Learning Nearly Decomposable Value Functions with Communication Minimization" (ICLR 2020)

NDQ: Learning Nearly Decomposable Value Functions with Communication Minimization Note This codebase accompanies paper Learning Nearly Decomposable Va

Tonghan Wang 69 Nov 26, 2022
Spectralformer: Rethinking hyperspectral image classification with transformers

Spectralformer: Rethinking hyperspectral image classification with transformers Danfeng Hong, Zhu Han, Jing Yao, Lianru Gao, Bing Zhang, Antonio Plaza

Danfeng Hong 102 Dec 29, 2022
HyDiff: Hybrid Differential Software Analysis

HyDiff: Hybrid Differential Software Analysis This repository provides the tool and the evaluation subjects for the paper HyDiff: Hybrid Differential

Yannic Noller 22 Oct 20, 2022
GenGNN: A Generic FPGA Framework for Graph Neural Network Acceleration

GenGNN: A Generic FPGA Framework for Graph Neural Network Acceleration Stefan Abi-Karam*, Yuqi He*, Rishov Sarkar*, Lakshmi Sathidevi, Zihang Qiao, Co

Sharc-Lab 19 Dec 15, 2022
Fedlearnๆ”ฏๆŒๅ‰ๆฒฟ็ฎ—ๆณ•็ ”ๅ‘็š„Pythonๅทฅๅ…ทๅบ“ | Fedlearn algorithm toolkit for researchers

FedLearn-algo Installation Development Environment Checklist python3 (3.6 or 3.7) is required. To configure and check the development environment is c

89 Nov 14, 2022
N-gram models- Unsmoothed, Laplace, Deleted Interpolation

N-gram models- Unsmoothed, Laplace, Deleted Interpolation

Ravika Nagpal 1 Jan 04, 2022
QICK: Quantum Instrumentation Control Kit

QICK: Quantum Instrumentation Control Kit The QICK is a kit of firmware and software to use the Xilinx RFSoC to control quantum systems. It consists o

81 Dec 15, 2022
Machine-in-the-Loop Rewriting for Creative Image Captioning

Machine-in-the-Loop Rewriting for Creative Image Captioning Data Annotated sources of data used in the paper: Data Source URL Mohammed et al. Link Gor

Vishakh P 6 Jul 24, 2022
Collection of NLP model explanations and accompanying analysis tools

Thermostat is a large collection of NLP model explanations and accompanying analysis tools. Combines explainability methods from the captum library wi

126 Nov 22, 2022
Codes to calculate solar-sensor zenith and azimuth angles directly from hyperspectral images collected by UAV. Works only for UAVs that have high resolution GNSS/IMU unit.

UAV Solar-Sensor Angle Calculation Table of Contents About The Project Built With Getting Started Prerequisites Installation Datasets Contributing Lic

Sourav Bhadra 1 Jan 15, 2022
Avatarify Python - Avatars for Zoom, Skype and other video-conferencing apps.

Avatarify Python - Avatars for Zoom, Skype and other video-conferencing apps.

Ali Aliev 15.3k Jan 05, 2023
Install alphafold on the local machine, get out of docker.

AlphaFold This package provides an implementation of the inference pipeline of AlphaFold v2.0. This is a completely new model that was entered in CASP

Kui Xu 73 Dec 13, 2022