ROBEL: Robotics Benchmarks for Learning

ROBEL is an open-source platform of cost-effective robots and associated reinforcement learning environments for benchmarking reinforcement learning in the real world. It provides Gym-compliant environments that easily run in both simulation (for rapid prototyping) and on real hardware. ROBEL robots are robust and scalable - they have facilitated over 14000 hours (as of Aug'19) of real-world training with various learning-based methods. Benchmarks using several learning-based methods are provided for ease of comparison and extensibility. Refer to ROBEL's webpage for full details.

Robot Platforms

ROBEL robots: D’Kitty (left) and D’Claw (middle and right)

ROBEL introduces two robots that are modular, extensible, easy to build, and are lower cost compared to many existing robotics research platforms:

1. D'Claw is a nine degree of freedom manipulation platform capable of performing dexterous manipulation.
2. D'Kitty is a twelve degree of freedom quadruped capable of agile locomotion.

Features

1. Gym Compliant -- ROBEL environments are fully Gym-compliant and can be used with any reinforcement learning library that interfaces with Gym environments.
2. Simulated backends -- ROBEL also includes simulated equivalents of the introduced benchmarks to facilitate prototyping and debugging needs. Simulation backend is provided by MuJoCo.
3. Hardware interface -- ROBEL is built using Dynamixel motors and communicates with the hardware device through the DynamixelSDK.
4. External tracking support -- For D'Kitty environments, external tracking is supported through OpenVR tracking.
5. Open-source design -- The hardware design and build instructions are fully open-sourced and are available for anyone to build their own robots.

Getting started

1. Hardware assembly

Please refer to the Hardware Guide for getting started with the ROBEL hardware platforms.

2. Software installation

2.1 MuJoCo

Download MuJoCo Pro 2.00 from the MuJoCo website. You should extract this to ~/.mujoco/mujoco200. Ensure your MuJoCo license key is placed at ~/.mujoco/mjkey.txt.

Add the following line to your ~/.bashrc (or equivalent) in order for mujoco_py to install properly:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HOME/.mujoco/mujoco200/bin

Run source ~/.bashrc afterwards.

2.2 ROBEL

ROBEL requires Python 3.5 or higher. You can install ROBEL by running:

pip install robel

We recommend doing this in a virtualenv or a Conda environment to avoid interfering with system dependencies or existing packages.

Alternatively, you can install directly from the repository by running:

git clone --recurse-submodules https://github.com/google-research/robel.git
pip install -e robel/

To run on hardware, additionally install the DynamixelSDK Python API:

pip install git+https://github.com/ROBOTIS-GIT/DynamixelSDK.git#subdirectory=python

2.3 Example Usage

import robel
import gym

# Create a simulation environment for the D'Claw turn task.
env = gym.make('DClawTurnFixed-v0')

# Create a hardware environment for the D'Claw turn task.
# `device_path` refers to the device port of the Dynamixel USB device.
# e.g. '/dev/ttyUSB0' for Linux, '/dev/tty.usbserial-*' for Mac OS.
env = gym.make('DClawTurnFixed-v0', device_path='/dev/ttyUSB0')

# Reset the environent and perform a random action.
env.reset()
env.step(env.action_space.sample())

Not specifying the device_path i.e. env = gym.make('DClawTurnFixed-v0') creates the simulated equivalent of the above hardware environment. The simulated and hardware environments have the same interface.

To interactively render a simulation environment, run:

python -m robel.scripts.rollout -e DClawTurnFixed-v0 --render
# Also try this with other tasks such as DKittyWalkFixed-v0

Benchmark Tasks

D'Claw

D'Kitty

Contributing

We designed ROBEL to be an easily extensible platform for new robots, tasks, and benchmarks. See CONTRIBUTING.md for a guide on how to contribute.

Citation

@misc{ahn2019robel,
    title={ROBEL: Robotics Benchmarks for Learning with Low-Cost Robots},
    author={Michael Ahn and Henry Zhu and Kristian Hartikainen and Hugo Ponte and Abhishek Gupta and Sergey Levine and Vikash Kumar},
    year={2019},
    eprint={1909.11639},
    archivePrefix={arXiv},
    primaryClass={cs.RO}
}

Disclaimer

This is not an official Google product.