Skild Steel

Skild Steel (also called Steel / STEEL — Scalable Throughput-optimized Engine for Embodied Learning) is an open GPU rigid-body physics solver from Skild AI, developed with NVIDIA, for the Newton engine used with Isaac Lab.

It integrates as newton.solvers.SolverSteel, targets heterogeneous multi-robot batches, and is released under Apache-2.0 with a WebGPU playground and code at https://github.com/skild-ai/steel . Launch materials report roughly 3.8× PhysX geometric-mean throughput and up to about 13× on mixed robot-task batches, on the order of ~3.9M environment steps per second.

Highlights:

• Official Skild post: https://skild.ai/blogs/steel ;
• Open GPU physics solver for Newton / Isaac Lab robot training;
• AKA Steel, STEEL, newton.solvers.SolverSteel;
• Heterogeneous multi-robot batches; Apache-2.0; WebGPU playground;
• Repo: https://github.com/skild-ai/steel .

Why it matters

Embodied learning throughput is often limited by the physics solver, especially when batches mix many robot types. Skild Steel attacks that bottleneck inside Newton for Isaac Lab: a GPU rigid-body solver tuned for scalable, heterogeneous robot training rather than single-env demos. The named speedups versus PhysX (about 3.8× geomean, up to ~13× on mixed batches, ~3.9M env steps/s as reported) are why robotics labs watching CoRL-adjacent launches care.

Open Apache-2.0 licensing, a public GitHub repo, and a WebGPU playground lower the barrier to try the solver without a closed vendor binary. Radar recorded the product as Skild Steel to avoid the generic word “Steel” colliding with unrelated hits.

Limits

Throughput multipliers depend on batch composition, GPU generation, and Newton/Isaac Lab versions—reproduce on your workloads before citing them in capacity plans. Rigid-body scope may not cover every soft-body or continuum need. Confirm current Newton integration path, CUDA/WebGPU support, and Apache-2.0 obligations in the repo. Do not treat unpublished memory footprints or private benchmark forks as facts.

1. Read https://skild.ai/blogs/steel and clone https://github.com/skild-ai/steel ;
2. Wire newton.solvers.SolverSteel into your Isaac Lab / Newton training loop;
3. Benchmark mixed multi-robot batches against your current PhysX (or other) baseline on the same GPU;
4. Check Apache-2.0 terms and Newton version compatibility before fleet training; re-measure steps/s on your task mix.