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

Naive usage⁠(opens in a new window) of the new version’s MjSimPool⁠(opens in a new window) interface shows a 400% speedup over the old, and still about 180% over an optimized and restricted usage pattern using Python’s multiprocessing package⁠(opens in a new window) 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.

We use the domain randomization⁠(opens in a new window) 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(opens in a new window) for an example of randomized texture generation.

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(opens in a new window) to Python using mujoco-py. If you have an HTC Vive VR setup, you can give try it using this example⁠(opens in a new window) (this support is considered experimental, but we’ve been using it internally for a while).