Skild AI is open-sourcing Steel, a GPU rigid-body physics solver built with NVIDIA for robot learning. Steel (Scalable Throughput-optimized Engine for Embodied Learning) runs inside the open-source Newton physics engine and is built to speed simulation in NVIDIA Isaac Lab. It is for researchers training manipulation policies who need many different robots and tasks in one batch. GPU simulators usually run fastest when every environment copies the same robot and the same task. Steel groups articulations with similar joint counts and assigns contact-solving work according to each environment, so a mixed batch of arms, hands, quadrupeds, and humanoids does not pad every world to the heaviest scene. In Skild's measurements on one NVIDIA RTX PRO 6000 Blackwell GPU, Steel delivered 3.8 times PhysX simulation throughput on a geometric mean across 20 robot-learning benchmarks, and was faster on 19 of the 20. On batches mixing different task types, peak throughput reached as high as 13 times PhysX. At a matched 61,440 environments in a 20-task mix, Skild reports 12 times the environment steps per second and about 62 percent less GPU memory, roughly 11 GiB versus 29 GiB. Those figures cover physics stepping only, not full policy-training speed. On the same single-task suite, Steel was about 1.3 times MuJoCo Warp. The code is Apache 2.0, available in Newton as SolverSteel, with a WebGPU build for live browser demos. Skild is a Pittsburgh company founded in 2023 by Deepak Pathak and Abhinav Gupta, both from Carnegie Mellon, with Pathak as CEO and Gupta as president. The company builds general-purpose robotic intelligence meant to control many bodies and tasks, and it relies on large-scale simulation to train across that range. Opening the internal solver gives other labs the tool Skild used for those mixed manipulation workloads.