Open Source Robotics

A curated collection of high-quality open source robotics projects, tools, and software to propel the robotics community forward.

Steel

Skild AI's GPU physics solver for simulating many different robots and tasks at once

License

Apache 2.0

Code: Apache-2.0, per the LICENSE file in the Steel repository; the solver itself is distributed in Newton, which is also Apache-2.0. Media and marks: the NOTICE file says the licence does not cover third-party assets, trademarks, the launch footage or its music. Browser port: steel-webgpu is Apache-2.0, and the robot models bundled with it keep their own BSD-3-Clause and MIT licences. Checked 2026-11-10 against LICENSE, NOTICE and the steel-webgpu LICENSE.md.

Overview

Steel is a GPU rigid-body physics solver from Skild AI, built with NVIDIA and released as open source on 9 November 2026 as part of the Newton physics engine, where it is available as SolverSteel. It is designed for robot learning in simulation when one batch mixes very different scenes: arms, dexterous hands, quadrupeds and humanoids, each with its own joints and contacts. It combines reduced-coordinate articulation dynamics with a projected Gauss-Seidel constraint solve, and sizes the work for each environment instead of padding every scene to the largest. Skild reports 3.8 times the simulation throughput of PhysX on average across 20 robot-learning tasks and up to 13 times on mixed batches, measured by Skild on one NVIDIA RTX PRO 6000 Blackwell GPU. A WebGPU port runs the solver in the browser, with a public playground. Steel is Apache-2.0 and is still marked experimental in Newton.

Steel (Scalable Throughput-optimized Engine for Embodied Learning) is a rigid-body dynamics solver written by Skild AI together with NVIDIA for Newton, the open-source GPU physics engine built on NVIDIA Warp that is a Linux Foundation project. The solver code lives in the Newton repository, where it was added on 15 October 2026 (pull request) and is used as newton.solvers.SolverSteel; the Steel repository holds the launch README, licence and video. Skild announced the open-source release on 9 November 2026 in a launch post, which says Steel is meant to speed up robot learning in Isaac Lab.

Steel targets a case that many GPU simulators are not built for: a batch of thousands of environments that are not copies of one another. A robot arm reaching in free space, a dexterous hand turning a cube and an arm working through a cluttered bin have different numbers of joints, constraints and contacts, and the contacts change at every step. Steel keeps the dynamics and constraints of each environment separate, groups articulations with similar joint counts for the GPU, and routes contact work to each environment according to what it currently needs, which reduces padding and wasted work. Each step builds and factors the joint-space mass matrix of every articulation and solves contacts, joint limits and other constraints by projected Gauss-Seidel, reusing that factorisation without assembling a dense matrix of contact interactions. The Newton documentation lists its features: prismatic, revolute, ball, fixed, D6 and free joints, joint drives and limits, mimic joints, closed kinematic loops, friction patches and sleeping of settled bodies.

Because it shares Newton's model and collision pipeline, Steel works alongside Newton's other solvers. The launch post shows it as the rigid side of coupled scenes with cables, cloth, granular sand and soft bodies, with hydroelastic contact, and with an experimental differentiable step that passes gradients through contact.

All speed figures are Skild's own, for simulation only rather than end-to-end training, measured from 7 to 9 November 2026 on one NVIDIA RTX PRO 6000 Blackwell GPU against PhysX from Isaac Sim 6.1 and MuJoCo Warp 3.17, each engine at its best batch size. Across 20 single-task benchmarks Steel reached 3.8 times PhysX's throughput as a geometric mean and was faster on 19 of the 20 tasks; against MuJoCo Warp it was about 1.3 times as fast. On batches mixing 5, 10, 15 or 20 task types, including humanoids, quadrupeds, arms and dexterous hands, it reached 9.2 to 13.0 times PhysX's peak throughput. With 20 task types and 61,440 environments for both engines, Steel ran 12.0 times as fast and used about 62% less GPU memory, about 11 GiB against 29 GiB.

Skild also released steel-webgpu, a port of the solver to WebGPU written in TypeScript and published on npm as @skild-ai/steel-webgpu. It loads scenes from URDF, MJCF and USD files, draws them with three.js or React Three Fiber, and runs in Chrome, Edge or Safari. The Steel playground uses it to run 24 scenes live in the browser, from Jenga and dominoes to a trebuchet, where you can drag objects and change the solver settings while it runs.

Why it matters

GPU simulation made reinforcement learning for walking robots practical: thousands of copies of one quadruped or humanoid practise in parallel, with domain randomization helping the result carry over to hardware. Manipulation is harder to scale that way, because useful training data needs many different tasks, objects and robot bodies, while many GPU simulators, including those behind Isaac Lab, are most efficient when every environment in a batch is the same. Steel is an open solver built for the mixed case, and it sits inside Newton next to the MuJoCo Warp solver, so researchers can switch to it without adopting a new physics stack and can read, test and change how it works. It is also Skild AI's first open-source release, a company that builds robot foundation models and uses simulation to train them, and the browser playground lets anyone try the solver without a GPU workstation. Whether its speed-ups hold in real training pipelines, and help close the sim-to-real gap, is still to be shown by people outside Skild.

Problem it solves

GPU robot simulators are fastest when they run thousands of copies of the same robot and task, but training general manipulation policies needs batches that mix different robots, objects and tasks, each with its own joints and changing contacts. Padding every environment to the largest one, or building dense matrices for complex articulations, wastes GPU time and memory. Steel sizes the solve for each environment, groups similar articulations and avoids dense contact matrices, so one GPU can step mixed batches of arms, hands, quadrupeds and humanoids with less wasted work.

Key takeaways

  • A rigid-body physics solver from Skild AI and NVIDIA, released on 9 November 2026 inside the Newton physics engine as SolverSteel
  • Built for batches that mix different robots and tasks: work is sized per environment and articulations are grouped by joint count
  • Reduced-coordinate articulation dynamics with a matrix-free projected Gauss-Seidel solve of contacts and constraints
  • Skild's own benchmarks on one RTX PRO 6000 Blackwell GPU: 3.8 times PhysX on 20 single tasks and 9.2 to 13.0 times on mixed batches
  • Couples with Newton's cloth, cable, soft-body and granular solvers, with an experimental differentiable step
  • A WebGPU port, steel-webgpu, runs the solver in the browser and powers a public playground
  • Apache-2.0, in both the Steel repository and Newton

Who it is for

  • Robot learning researchers training manipulation policies in simulation across many tasks and robots
  • Isaac Lab and Newton users looking for a faster solver for articulated robots with many contacts
  • Physics simulation developers who want to study or extend an open GPU contact solver
  • Web developers who want interactive rigid-body physics with robot models in the browser

Limitations

  • All speed and memory figures come from Skild's own measurements on one NVIDIA RTX PRO 6000 Blackwell GPU and have not been reproduced by others
  • The figures cover physics stepping only, without observations, rewards, resets, rendering or policy training, so end-to-end training gains will be smaller
  • The PhysX comparison uses Isaac Sim 6.1; Skild notes that PhysX from Isaac Sim 6.0 ran the mixed batches up to 2.7 times faster, so the gap is smaller against that version
  • The single-task results use an option, reuse_kinematics, that is opt-in in Newton 1.8 and turned off when joint velocity limits are on
  • Newton marks SolverSteel as experimental, so its options and attributes may change without the normal deprecation period
  • The Steel repository holds only a README, licence and launch video; the code is in Newton, which is itself in release candidates for version 1.8
  • Steel solves rigid bodies only; cloth, cables, soft bodies and sand come from Newton's other solvers through coupling
  • Released on 9 November 2026, so it has had almost no use outside Skild and NVIDIA

This website requires JavaScript to function properly. Please enable JavaScript to explore our collection of open source robotics projects.