Skild AI and NVIDIA release Steel, a physics solver Skild says runs robot-learning simulation 3.8 times faster than PhysX

Skild.ai and NVIDIA introduced Steel, a GPU-accelerated rigid-body dynamics solver for the open-source Newton Physics Engine used for robot learning in NVIDIA Isaac Lab, Skild.ai said on 10 November. Skild.ai says its own simulation-only tests across 20 robot-learning benchmarks found Steel delivered 3.8 times the geometric mean throughput of PhysX and ran faster on 19 of the 20.

On a mixture of the 20 tasks on one NVIDIA RTX PRO 6000 Blackwell GPU, Steel processed 3.90 million environment steps per second, at 245,760 environments, against 300,000 for PhysX at 61,440 environments. At a matched batch size of 61,440 environments, Steel reached 3.61 million steps per second and used about 11 GiB of GPU memory, while PhysX used about 29 GiB.

Steel groups robot articulations with similar joint counts and assigns contact-solving work according to each environment's current demands. The solver is available in Newton as newton.solvers.SolverSteel and is designed for robot training simulations in NVIDIA Isaac Lab.