# Nvidia Warp and MJWarp enable up to 2,048 parallel robot simulations

Digest AI · Agents & Tools · published 2026-09-23T18:41:40Z

Canonical: https://digestai.news/story/nvidia-warp-and-mjwarp-enable-up-to-2-048-parallel-robot-simulations

## Summary

It walks through validating a single world, converting it with mjw.putmodel and mjw.makedata, and then scaling the batch by tiling the initial state across worlds. The guide details required parameters such as nconmax=128 and njmax=300 for the default profile and nconmax=256 and njmax=500 for the optional reBot variant, and explains how to capture a CUDA graph for repeated stepping.

Key steps include installing warp‑lang (≥ 1.15 for optional deterministic execution) and mujoco‑warp, using wp.ScopedCapture to record the step graph, and timing the throughput with a warm‑up of ten launches followed by 200 timed launches. The article also lists batch‑size experiments (1, 64, 1024, 2048, 8192 worlds) and provides command‑line examples for running the SO‑101 pick‑and‑place task at 50 Hz control frames with 10 physics substeps (0.002 s timestep).

## Key points

- MJWarp can batch up to 2,048 independent worlds on a single NVIDIA GPU.
- Warp 1.15 adds optional deterministic execution for reproducible simulations.
- Default batch limits are nconmax=128 and njmax=300; reBot variant uses nconmax=256 and njmax=500.

## Why it matters

Scaling robot simulation to thousands of parallel worlds speeds up reinforcement‑learning data collection and design optimization without changing model code.

## Sources

1. [How to Use NVIDIA Warp and MjWarp to Accelerate Robotics Simulation and Learning Workflows](https://huggingface.co/blog/nvidia/how-to-use-nvidia-warp-and-mjwarp) (Hugging Face, 2026-09-23, primary source)

## Cite

Digest AI, "Nvidia Warp and MJWarp enable up to 2,048 parallel robot simulations", 23 September 2026, https://digestai.news/story/nvidia-warp-and-mjwarp-enable-up-to-2-048-parallel-robot-simulations

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/nvidia-warp-and-mjwarp-enable-up-to-2-048-parallel-robot-simulations.json