How to Use NVIDIA Warp and MjWarp to Accelerate Robotics Simulation and Learning Workflows

NVIDIA's Warp and MjWarp enable accelerating robotics simulation and learning workflows by leveraging GPU parallelism. While classic MuJoCo offers fast CPU-based simulation, as workloads grow, the focus shifts to running many worlds simultaneously. GPU acceleration allows advancing multiple worlds in large batches while maintaining close simulation and learning data proximity.

MuJoCo Warp (MJWarp) connects Python to GPU simulation, with MuJoCo loading and compiling the MJCF model, and MJWarp implementing physics in NVIDIA Warp, which compiles CUDA kernels for state advancement on NVIDIA GPUs. The article examines the transition of a SO-101 follower arm from a familiar MuJoCo workflow to 2,048 parallel MJWarp environments.

Warp is a Python framework for writing high-performance, GPU-accelerated kernels. Developers author statically typed kernels in Python, which are then compiled for CPU or CUDA execution. The kernel language is a performance-oriented subset of Python, while ordinary Python handles configuration, allocation, and launch orchestration. Warp kernels are differentiable, allowing teams to build differentiable geometry, CFD, and custom physics.

In terms of robot simulation, each world represents an independent copy of the scene and its state. MuJoCo and MJWarp can run the same compatible robot and task, but they differ in how they organize the work. MuJoCo is suitable for developing and inspecting one or a few CPU worlds, while MJWarp, built on NVIDIA Warp, can advance hundreds or thousands of worlds at once, improving aggregate throughput.

To migrate to MJWarp, use mjw.make_data() for default/fresh state or mjw.put_data() for the exact initialized MuJoCo state. Allocate batched resources by defining CUDA graph capture, size parameters (nconmax, naconmax, njmax), and tuning with mjwarp-testspeed --measure_alloc. Additional tuning considerations include solver iteration limits, mesh and CCD settings, and nccdmax/naccdmax.

The article provides an SO-101 arm, a table, and two cubes to stack as an example scene. The task is to grasp the red 44 mm cube and stack it on the blue cube. The robot and scene are used for MJWarp validation.