Skild AI has released Steel, a GPU-accelerated rigid-body physics solver for the open-source Newton physics engine, built with Nvidia for robot learning in Isaac Lab. Skild AI published the technical write-up this month, with benchmarks run on November 7 to 9. The name stands for Scalable Throughput-optimized Engine for Embodied Learning. The code is on GitHub at skild-ai/steel, and the solver ships in Newton as newton.solvers.SolverSteel.
Steel is designed to simulate mixed batches of different robots, tasks and scenes on one GPU. It groups robots by joint count, solves contacts separately for each environment, and combines reduced-coordinate articulation dynamics with projected Gauss-Seidel constraint solving.
On a single Nvidia RTX PRO 6000 Blackwell GPU, Skild AI reports that Steel averages 3.8 times the throughput of PhysX across a 20-task suite, and is faster on 19 of the 20 tasks. With mixed batches the gap widens: 9.2 times PhysX with 5 task types and 13.0 times with 20. In one matched run with 20 task types and 61,440 environments, Steel reached 3.61 million environment steps per second against 300,000 for PhysX, using about 11 GiB of GPU memory against 29 GiB. Skild AI says Steel is about 1.3 times faster than MuJoCo Warp on the same 20 tasks.
Skild AI notes that these figures measure simulation throughput only, not end-to-end policy training speed. The single-task results used an opt-in option in the Newton 1.8 release candidate, and PhysX was run from Isaac Sim 6.1. A WebGPU version, steel-webgpu, runs 24 demo scenes in a browser.
Simulation throughput often limits how fast robot policies can be trained with reinforcement learning, especially for generalist policies that train on many tasks at once. Steel runs in Newton, which Nvidia's Isaac Lab 3.0 now lets users choose instead of PhysX, so teams already on Isaac Lab can test the solver without changing their simulation stack.