MuJoCo Warp (MJWarp), built on NVIDIA Warp, allows compatible MuJoCo models to run at GPU scale, enabling the advancement of hundreds or thousands of independent simulation worlds in a single call. This architecture shifts the focus from minimizing latency for one environment to improving aggregate throughput, which favors reinforcement learning and large-scale sampling.

  • MJWarp implements MuJoCo’s physics pipeline using NVIDIA Warp, compiling CUDA kernels to advance simulation states on NVIDIA GPUs.
  • The framework supports explicit parallel work, device arrays, and composable kernel launches, with optional differentiability and deterministic execution modes.
  • Performance tuning involves CUDA graph capture for repeated steps and tight sizing of contact and constraint buffers like nconmax and njmax.
  • Migration from CPU MuJoCo requires matching simulation rates and validating results against a baseline, such as an SO-101 arm stacking cubes.

This approach helps users scale robotics simulation workflows by leveraging GPU parallelism for efficient experience collection in training pipelines.