Skild AI and Nvidia's Steel solver hits 13x PhysX throughput for robot learning

lukas_m_ziegler · x · 2026-11-17

Skild AI and NVIDIA released Steel, a GPU-accelerated rigid-body dynamics solver for the open-source Newton physics engine in Isaac Lab: 3.8x PhysX geometric-mean throughput across 20 robot-learning benchmarks, 9.2–13.0x peak on heterogeneous batches mixing 5–20 task types, and 62% less GPU memory at matched batch size. Also: Runway open-sourced Praxis-1 (Apache 2.0), a 27B MoE diffusion transformer topping RoboLab-120 at 46.2% success vs FLUX 3 Action 42.9%, Cosmos3 Nano 36.8%, π0.5 28.0%, GR00T N1.6 7.3%, with commercial-use weights and code.

Original post →

More from Embodied

Embodied channel →