Sakhanda Wire
NVDA $153.64 -1.60% MSFT $506.98 -2.13% GOOGL $279.95 -1.33% META $620.09 -1.01% AMZN $256.57 -0.46%
← Back to the news

NVIDIA Warp brings MuJoCo robotics simulation to batched GPU execution

A new NVIDIA guide explains how MuJoCo robotics workflows can be moved from CPU execution to MuJoCo Warp, or MJWarp, which implements compatible MuJoCo physics with NVIDIA Warp on NVIDIA GPUs. The walkthrough uses an SO-101 follower arm in a pick-and-place scene containing a table and two cubes. It focuses on preparing and validating the simulation environment rather than training a policy.

The central change is replacing a single MuJoCo model and data object with device-resident, batched resources. After the model is uploaded, one MJWarp step can advance many independent worlds; the guide scales the example to as many as 2,048 environments. This approach targets aggregate throughput for reinforcement learning and large-scale sampling, rather than necessarily reducing the latency of one simulation step.

The article also introduces Warp kernels, which can run on CPUs or CUDA devices and support features including automatic differentiation, PyTorch and JAX interoperability, and CUDA Graphs. It advises sizing contact and constraint buffers carefully and distinguishing latency from total world-steps per second when measuring performance. MJWarp can be used with mjlab, MuJoCo Playground through MJX, and future Newton and Isaac Lab integration layers.

Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.

Originally published by Hugging Face on

Read the original on Hugging Face ↗

The reporting and the image belong to Hugging Face and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news