Skild AI and NVIDIA Open-Source Steel, a Robot Physics Solver for Newton That Hits 13x PhysX Peak Throughput on Mixed 20-Task Batches With About 62% Less GPU Memory

Background and Pain Point

GPU-parallel simulation has driven big gains in locomotion because thousands of copies of the same robot and task can train at once. Manipulation is different: it needs huge numbers of diverse, contact-rich scenes, and robots differ widely in joint and contact counts. Skild AI argues that now that agents can write those tasks at scale, the bottleneck is simulating them together. Many engines are most efficient on homogeneous batches and waste compute and memory on padding and dense matrices when batches are heterogeneous.

Architecture and Mechanism

Steel (Scalable Throughput-optimized Engine for Embodied Learning) is a rigid-body solver inside the Newton physics engine. It organizes dynamics and constraints per environment, groups articulations with similar joint counts, and routes contact-solving work by each environment's current load. Under the hood it combines reduced-coordinate articulation dynamics with projected Gauss-Seidel constraint solving and reuses each articulation's mass-matrix factorization without building a dense pairwise constraint matrix. It shares Newton's model and collision stack, couples with VBD cloth/cable and MPM sand solvers, and offers experimental differentiable simulation.

Benchmarks

All vendor figures, simulation-only, one RTX PRO 6000 Blackwell GPU: on mixed batches of 5/10/15/20 task types Steel reaches 9.2x to 13.0x PhysX peak throughput; at 20 types it hits 3.90M steps/s vs 300K for PhysX. At a matched 61,440 environments it is 12x faster and uses about 11 GiB vs 29 GiB. Across 20 single-task benchmarks it averages 3.8x PhysX (MuJoCo Warp: 2.9x) and leads on 19. The comparison uses Isaac Sim 6.1; Skild notes PhysX on 6.0 ran these mixtures up to about 2.7x faster. No third-party reproduction yet.

Getting Started

Code is at GitHub skild-ai/steel (Apache 2.0) and in Newton as newton.solvers.SolverSteel, aimed at Isaac Lab training. The single-task results used reuse_kinematics=True, which is opt-in in Newton 1.8. To try it first, open the WebGPU demo in the blog and drag objects or tune solver settings in the browser. Teams doing manipulation RL that need many robot types in one batch benefit most.