Hugging Face and NVIDIA detail workflow for accelerating robotics simulation with MuJoCo Warp

Hugging Face and NVIDIA published a joint technical guide describing how to use MuJoCo Warp (MJWarp) to accelerate robotics simulation and learning workflows on NVIDIA GPUs. The guide details a workflow for migrating compatible MuJoCo models from CPU-based simulation to MJWarp, a GPU-accelerated backend built on NVIDIA Warp.

MJWarp enables users to run hundreds or thousands of independent simulation worlds in parallel on a single GPU, with the guide demonstrating a batch of up to 2,048 parallel environments based on a MuJoCo SO-101 follower arm scene. The workflow emphasizes validating correctness, forming batches, and measuring throughput rather than optimizing per-environment latency.

NVIDIA Warp is presented as a Python framework for writing high-performance, GPU-accelerated kernels that compile to CUDA or CPU code. Warp kernels are statically typed, support autodiff via wp.Tape, and provide interoperability with PyTorch and JAX through DLPack-style sharing. The framework also supports deterministic execution modes introduced in Warp version 1.15, which trade some performance for reproducible ordering in simulation and validation contexts.

The guide walks through a minimal Warp kernel example that advances point positions under gravity, illustrating explicit parallel work via wp.tid(), device-resident arrays, and composable kernel launches that can be captured into CUDA graphs to reduce dispatch overhead. It clarifies that differentiability and determinism are capabilities of Warp kernels, not guaranteed properties of an entire MJWarp rollout, and points readers to Warp documentation for further details.

The authors position MJWarp as complementary to existing MuJoCo workflows: MuJoCo remains suitable for developing and inspecting one or a few CPU-based worlds, while MJWarp targets large-batch simulation for reinforcement learning and high-throughput sampling. The guide is part of a broader “State of Simulation for Physical AI” series, with subsequent installments planned to cover integration layers in Newton and Isaac Lab.