Classic MuJoCo simulates robots on CPUs and parallelizes sampling across cores, but learning workloads increasingly need many worlds at once. NVIDIA Warp and MJWarp address this by moving compatible MuJoCo models to GPU-accelerated simulation. The source article demonstrates the transition using an SO-101 follower arm, scaling from a familiar MuJoCo workflow to as many as 2,048 parallel MJWarp environments.
Warp is a Python framework for writing GPU kernels that compile to CUDA. It offers three pillars: performance through JIT compilation and kernel fusion, ease of use with pure Python authoring and built-in primitives, and capability via differentiable kernels and DLPack-style interop with PyTorch and JAX. The article includes a simple gravity-integration kernel to illustrate how one logical thread handles one point, so the same code scales from two points to millions.
The authors note that MJWarp is not the only path. For single-robot MPC or teleop, classic MuJoCo on CPU remains the choice; for JAX training recipes, MuJoCo Playground or MJX with impl='warp' may fit better. MJWarp is positioned for maximum throughput on raw MuJoCo physics, while the upcoming Newton post in the series will cover multi-solver integration with Isaac Lab. The article also flags that Warp's differentiability and deterministic execution are capabilities, not guarantees for an entire MJWarp rollout. Since there is only one source, no contrasting views are available. The piece is the second in the State of Simulation for Physical AI series and does not train a policy, focusing instead on preparing and scaling the simulation{