NVIDIA Warp guide scales one robot arm to 2,048 environments
Victoria Shinder2 min readNVIDIA and Hugging Face published the second part of their State of Simulation for Physical AI series on 23 September 2026, written by Johnny Nunez Cano: https://huggingface.co/blog/nvidia/how-to-use-nvidia-warp-and-mjwarp. It walks an SO-101 follower arm out of a single-world MuJoCo workflow and into as many as 2,048 parallel environments.
Two pieces do the work. Warp is a Python framework for GPU kernels: developers write statically typed kernels in Python and it compiles them for CPU or CUDA, building and caching a native module on first launch and reusing it afterwards. MJWarp is MuJoCo's physics pipeline implemented in Warp, holding the model and a batch of independent states on the GPU, where a single step call advances the whole batch.
The article is explicit that MJWarp's value is not a faster step for one world. It is the ability to advance hundreds or thousands of worlds together.
What it means
The interesting claim here is about the unit of work, not about speed. A simulator that is faster per world makes the same experiment finish sooner. A simulator whose unit is a batch makes a different experiment possible: hundreds of slightly different initial poses advanced in lockstep, which is the shape reinforcement learning actually wants. That is a change in what you can ask, and it is why the piece keeps saying the same model file runs on both.
And it explains why the guidance is a decision table rather than a benchmark. The article's own shortcut sends single-robot control and teleoperation to MuJoCo on the CPU and maximum throughput to MJWarp - because for one world, the batched engine has nothing to batch.
⚠️ Read it for what it is: a staging guide, not a training run. The piece says plainly that it prepares and scales the environment and does not train a policy, leaving the integration layers to later instalments. No performance figure of our own is reported here.