Skip to content
Skild AISkild AISteel

Steel

Steel (Scalable Throughput-optimized Engine for Embodied Learning) is a GPU solver for articulated robots and rigid bodies in Newton. Robots in joint coordinates, contacts and constraints in a matrix-free projected Gauss-Seidel sweep, and every world of a batch scheduled on its own.
Quick startTutorialsConfiguration reference
StartWhat Steel is for, which Newton build has which feature, installation and a first simulation.
TutorialsOne short script per feature, each with a 3D replay of a GPU recording of that exact script.
ConceptsThe step pipeline, the matrix-free sweep, friction patches, capacities, sleeping and determinism, with interactive demos.
GuidesTuning, porting from PhysX, domain randomization, observables, coupling and troubleshooting.
ReferenceEvery constructor option, steel: attribute, model input and method, generated from the solver's source.
ChangelogSteel changes in Newton and on the integration branch, newest first.

Examples

Section titled “Examples”

Scenes from the examples shelf, each run with its Steel preset: manipulation, contact-rich stacks, loop-closing linkages, cloth coupling and many worlds per GPU.

  • Franka shirt fold stack

    A Franka folds and stacks a T-shirt with two-way cloth coupling.

  • Jenga

    Lift a loose block and add it to the top of the tower.

  • Box turn

    Two arms pinch a box between their fingertips, flip it and turn it.

  • Domino spiral

    Dominoes along a spiral, the first one tilted to start the chain.

  • Card house

    Four levels of playing cards held up by friction alone.

  • Hoberman sphere

    A Hoberman sphere of 120 bars and 240 hinges opens and closes from one driven hub.

  • Robot panda hydro

    A Franka Panda picks and places with hydroelastic contacts and inverse kinematics.

  • Basic urdf

    A quadruped articulation loaded from URDF.

See all 81 examples →

Steel lives in Newton. This site, its tutorials and the examples are in skild-ai/steel; contributions are welcome, see Contributing. Throughput benchmarks are in the launch post.

Steel docs and tutorials are licensed under Apache-2.0. Third-party notices. "Skild", "Skild AI" and the Skild logo are trademarks of Skild AI and are not licensed under Apache-2.0.