Skild AI Unveils ‘Steel,’ a Physics Simulator for Handling Different Robots and Tasks Simultaneously

On Nov. 7, 2026, Skild AI and NVIDIA jointly announced Steel, a GPU rigid-body dynamics solver for the open-source physics engine Newton. Designed to accelerate robot training in NVIDIA Isaac Lab, Steel outperformed PhysX on 19 of 20 individual tasks, with a geometric mean throughput 3.8 times higher. In evaluations mixing different tasks, it recorded throughput 9.2 to 13.0 times PhysX’s peak.

Steel’s measurements showed 12.0 times the throughput of PhysX and about 62% lower GPU memory use at the same batch size, even when 20 different tasks were mixed. The results are relevant to researchers building simulations that handle multiple robot structures or tasks simultaneously. However, the figures are from single-GPU simulation measurements and do not demonstrate improvements in end-to-end training or real-world robot performance.

A notable aspect of Steel’s evaluation is that it demonstrated faster performance not only when replicating a single robot or task at scale, but also when handling different mechanisms and tasks within the same batch. The 13.0x figure for 20 mixed tasks compares the best batch size for each engine; at the same 61,440 environments, the advantage was 12.0x. The comparison conditions should be distinguished, though the results show throughput and memory-use differences even at the same batch size. This can be viewed as a design achievement for computing heterogeneous environments simultaneously. Our report on NVIDIA presents research on robots learning precision tasks through simulation at CoRL covered gear alignment, cable connections and small-part assembly. Steel focuses not on individual precision tasks, but on the computational infrastructure for processing robots and environments with different numbers of joints and contact states together. Both involve using simulation for robot learning, but Steel’s measurements alone do not establish that training time for a specific task or real-world success rates will improve. Structurally, Steel shares Newton’s model representation and collision-handling infrastructure, and is intended to integrate with training in Isaac Lab. It points toward handling different experiences within the same computational framework rather than preparing separate physics simulation environments for each robot or task. Skild AI says that the diversity of simulation experiences, not just their volume, matters for general-purpose policies. However, the figures released are for simulation throughput, not end-to-end training speed. Future evaluation will focus on how training with Steel affects policy performance and transfer to real robots, and how grounding and validation with real-world data are combined.

According to Skild AI, Steel’s throughput figures are for simulation only, measured on a single NVIDIA RTX PRO 6000 Blackwell GPU, with each engine evaluated at its best-performing batch size. In an evaluation mixing 20 tasks, Steel ran 3.9 million environment steps per second across 245,760 environments, while PhysX ran 300,000 steps per second across 61,440 environments. At the same 61,440 environments, Steel ran 3.61 million steps per second, 12.0 times PhysX’s rate, while using about 11 GiB of GPU memory compared with about 29 GiB, or about 62% less. In evaluations mixing 5, 10, 15 and 20 tasks, Steel’s throughput compared with PhysX’s peak was 9.2x, 10.4x, 12.5x and 13.0x, respectively. Steel groups mechanisms with similar numbers of joints and routes processing according to each environment’s contact state. Its design combines reduced-coordinate joint dynamics with constraint solving using the Projected Gauss-Seidel method. Steel is being made available through Newton’s open-source ecosystem and is intended to integrate with Isaac Lab.