Robotics / news brief
Skild details Steel benchmarks for mixed robot simulations
- Event
- Published

Skild AI published benchmarks for Steel, its experimental rigid-body solver inside the open-source Newton engine. Across 20 robot-learning tasks, Skild reports 3.8 times PhysX's geometric-mean simulation throughput and up to 13 times peak throughput on heterogeneous batches. Steel first appeared in Newton 1.8 RC1 on November 4, so today's new fact is the benchmark disclosure, not the initial code release; no independent reproduction was found.
Why it matters
Faster mixed-task simulation could diversify robot training data, but the reported gains are solver throughput under Skild's configurations, not end-to-end learning speed.
Sources
- Steel: Scaling Simulation Across Robots, Tasks, and Worlds (external site)Skild AI · primary source ·
- Skild AI Open-Sources Steel Physics Solver for Robot Training (external site)The Defiant Frontier · coverage ·