Robotics AI startup Skild AI released Steel, its internal physics solver, as open-source software on November 9, 2026, in partnership with Nvidia, according to the company's official blog post. Steel — short for Scalable Throughput-optimized Engine for Embodied Learning — is a GPU-accelerated rigid-body dynamics solver built on Nvidia's open-source Newton physics engine, designed to speed up the simulated training that robots rely on before they ever touch the real world.
In testing across 20 robot-learning benchmarks, Steel delivered 3.8 times the throughput of Nvidia's PhysX engine on uniform workloads, and between 9.2 and 13.0 times PhysX's peak throughput when running mixed batches of five to twenty different task types at once. In a matched-batch comparison at 61,440 simulated environments, Steel hit 12 times PhysX's throughput while using roughly 62% less GPU memory. Skild AI released the code under an Apache license and contributed it directly to the open-source Newton ecosystem, where it's available as newton.solvers.SolverSteel.
Why does simulation speed matter for robots?
Training a robot to grasp objects, use tools or complete multi-step tasks typically starts in simulation, where an AI system can attempt a task millions of times without breaking real hardware. The catch, known in robotics as the "sim2real gap," is that skills learned in an oversimplified simulated world often fail when transferred to a messy real one. Closing that gap usually means running simulations that are both more realistic and more varied — which costs more computing power.
Steel's architecture targets exactly that trade-off. It groups robot joints by configuration for efficient processing on Nvidia GPUs and routes contact-calculation work based on what each simulated environment actually needs at a given moment, rather than applying uniform computation across every scenario. That matters most in "heterogeneous" batches — training runs that mix different robots and different tasks simultaneously, which is closer to how real-world robot fleets need to learn than running the same single task over and over.
Why open-source it instead of keeping it proprietary?
Skild AI, co-founded by Carnegie Mellon researcher Deepak Pathak, builds general-purpose robot foundation models and has previously kept much of its technology closed. Releasing Steel as open infrastructure, integrated into Nvidia's community-maintained Newton engine, lets outside researchers inspect, reproduce and extend the solver rather than treating it as a Skild AI product.
That strategy mirrors a broader pattern in physical AI this year: companies compete on proprietary foundation models while open-sourcing the simulation and data infrastructure underneath them, on the theory that a faster, shared simulation layer benefits everyone building on Newton and Isaac Lab — including Skild AI's own future models. The release arrived alongside Nvidia's broader simulation push at the Conference on Robot Learning, where Nvidia and partners showcased a wave of tools aimed at making simulated training closer to real-world deployment.