Skild AI Open-Sources Steel, Its Physics Simulation Engine for Robotics Built with Nvidia
Skild AI has open-sourced Steel, a GPU-accelerated rigid-body dynamics solver developed alongside Nvidia and integrated into the Newton engine. According to its benchmarks, it simulates 3.8x faster than PhysX on average across 20 tasks, and up to 13x faster when mixing diverse robots and tasks.
To learn how to manipulate objects, a robot must train millions of times, and these trials happen first in simulation. Skild AI, one of robotics’ most high-profile startups, open-sourced Steel on November 9. The physics simulation engine, which it had been using internally, was developed alongside Nvidia. It joins Newton, the open-source physics engine spearheaded primarily by Nvidia, and integrates with Isaac Lab, Nvidia’s robot training framework.
Simulating Different Robots in the Same Batch
GPU-based simulators are remarkably efficient when cloning the exact same scene thousands of times: the same robot, the same task, with minor variations. That approach drove major breakthroughs in locomotion. Manipulation, however, is much more demanding: a robotic arm moving through empty space, a dexterous hand rotating a cube, and an arm rummaging through a cluttered bin do not share the same joint counts or contact dynamics. Mixing these setups forces conventional engines to “pad” simpler scenes to match the heaviest ones, wasting both compute and memory.
According to Skild’s blog post, Steel schedules workloads on an environment-by-environment basis, buckets robots with similar joint counts together, and sizes contact resolution dynamically based on each scene’s needs at every step. Under the hood, it pairs Featherstone’s articulated body dynamics with contact resolution using projected Gauss-Seidel, avoiding the construction of a large constraint interaction matrix.
Claimed Gains of 3.8x to 13x
The figures released by Skild measure simulation throughput alone on an RTX PRO 6000 Blackwell GPU. Across 20 standard robot learning tasks, Steel averages 3.8x the throughput of PhysX—Nvidia’s legacy engine—and roughly 1.3x that of MuJoCo Warp; it outperforms PhysX on 19 out of 20 tasks. The gap widens substantially on heterogeneous batches: by combining 20 task types (humanoids, quadrupeds, arms, hands), Steel tops out at 3.9 million steps per second, or 13x the peak throughput measured for PhysX. At identical batch sizes, it remains 12x faster while consuming about 62% less GPU memory.
Skild itself points out an important caveat: Isaac Sim 6.0 ran these heterogeneous mixes using PhysX up to 2.7x faster than version 6.1, which was used for this benchmark. Consequently, the exact speedup ratios depend on the chosen baseline.
What Is Available
The solver is accessible within Newton under the identifier newton.solvers.SolverSteel, currently flagged as experimental, meaning its configuration options may change without notice. The code is licensed under Apache 2.0 and hosted in the skild-ai/steel repository, which also links to a WebGPU version capable of running interactive demo scenes right in the browser. Steel can also serve as the rigid-body solver in Newton’s coupled multiphysics simulations (cables, cloth, sand, soft bodies) and offers an experimental differentiable stepping mode.
Our take
Skild’s bet is obvious: building general-purpose robots requires more than simulating at scale—it demands simulating diversity, which is precisely where today’s engines squander the most compute. Open-sourcing Steel within Newton rather than keeping it in an in-house tool gives it a real shot at becoming an industry standard. It also benefits Nvidia, which is pushing Newton and Isaac Lab against MuJoCo. Two caveats remain: all benchmarks come directly from Skild, using a PhysX baseline that isn’t the fastest version available, and higher simulation throughput alone does not guarantee better-trained robots. Early feedback from labs adopting the engine will show whether these gains hold up beyond Skild’s benchmark suite.
Sources
- Skild AI — Steel: Scaling Simulation Across Robots, Tasks, and Worlds (November 2026) Primary source
- GitHub — skild-ai/steel Primary source
- Newton Physics — newton.solvers.SolverSteel documentation Primary source
- GitHub — newton-physics/newton, Newton v1.8.0rc1 release Primary source
- Skild AI on X — Steel open-source announcement (November 9, 2026) Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
