Faster physics in Python

Faster physics in Python | OpenAI

June 28, 2017

Release

Faster physics in Python

We’re open-sourcing a high-performance Python library for robotic simulation using the MuJoCo engine, developed over our past year of robotics research.

View code View docs

Loading…

Share

This library is one of our core tools for deep learning robotics research⁠, which we’ve now released as a major version of mujoco-py⁠, our Python 3 bindings for MuJoCo. mujoco-py 1.50.1.0⁠ brings a number of new capabilities and significant performance boosts. New features include:

• Efficient handling of parallel simulations
• GPU-accelerated headless 3D rendering
• Direct access to MuJoCo functions and data structures
• Support for all MuJoCo 1.50 features⁠ like its improved contact solver

Batched simulation

Many methods in trajectory optimization and reinforcement learning (like LQR⁠, PI2⁠, and TRPO⁠) benefit from being able to run multiple simulations in parallel. mujoco-py uses data parallelism through OpenMP⁠ and direct-access memory management through Cython⁠ and NumPy⁠ to make batched simulation more efficient.

Naive usage⁠ of the new version’s MjSimPool⁠ interface shows a 400% speedup over the old, and still about 180% over an optimized and restricted usage pattern using Python’s multiprocessing package⁠ to gain the same level of parallelism. The majority of the speedup comes from reduced access times to the various MuJoCo data structures. Check out[examples/simpool.py](https://github.com/openai/mujoco-py/blob/master/examples/simpool.py) for a tour of MjSimPool.

High performance texture randomization

We use the domain randomization⁠ technique across many projects at OpenAI. The latest version of mujoco-py supports headless GPU rendering; this yields a speedup of ~40x compared to CPU-based rendering, letting us generate hundreds of frames per second of synthetic image data. In the above (slowed down) animation we use this to vary the textures of one of our robots, which helps it identify its body when we transfer it from the simulator to reality. Check out examples/disco_fetch.py⁠ for an example of randomized texture generation.

Virtual Reality with mujoco-py

The API exposed by mujoco-py is sufficient to enable Virtual Reality interaction without any extra C++ code. We ported MuJoCo’s C++ VR example⁠ to Python using mujoco-py. If you have an HTC Vive VR setup, you can give try it using this example⁠(this support is considered experimental, but we’ve been using it internally for a while).

API and usage

The simplest way to get started with mujoco-py is with the MjSim class⁠. It is a wrapper around the simulation model and data, and lets you to easily step the simulation and render images from camera sensors. Here’s a simple example:

Python

1from mujoco_py import load_model_from_path, MjSim23model = load_model_from_path("xmls/tosser.xml") 45sim = MjSim(model)6sim.step()7print(sim.data.qpos)8# => [ -1.074e-05 1.043e-04 -3.923e-05 0.000e+00 0.000e+00]

For advanced users, we provide a number of lower-level interfaces for accessing the innards of the MuJoCo C structs and functions directly. Refer to the README⁠ and the full documentation⁠ to learn more.

• Community & Collaboration
• Simulated Environments
• Software & Engineering
• Robotics

Authors

Jonas Schneider, Peter Welinder, Alex Ray, Jonathan Ho, Wojciech Zaremba

Related articles

View all

Introducing WhisperReleaseSep 21, 2022

Techniques for training large neural networksPublicationJun 9, 2022

Introducing Triton: Open-source GPU programming for neural networksReleaseJul 28, 2021