Newton: Open-Source GPU Physics for Robot Simulation | Physical AI Day at SIGGRAPH 2026

In this talk

1. 0:07

   Why Newton was built

   Miles describes simulation as a way to generate scalable synthetic data with physical interactions such as contact, friction, and forces. Newton was designed for robotics’ demands for fidelity, robustness, extensibility, GPU scale, and integration with machine-learning frameworks. He says NVIDIA, Google DeepMind, and Disney Research began the project together, with Lightwheel and Toyota Research Institute later joining.

2. 6:11

   Architecture and rigid-body simulation

   Newton separates physical models, states, controls, contacts, and solvers, allowing different numerical methods to share the same framework. It supports both joint-level and Cartesian representations, several solver types, and contact methods including hydroelastic contact. Miles cites a Skild AI assembly task as a sim-to-real example and a tactile grasp policy trained from simulated sensor readings.

3. 15:31

   Deformables, coupling, and performance

   Newton represents deformables as particles, rods, thin shells, or tetrahedral solids, with examples including snow, cables, cloth, and a sponge handled by a robot. Miles also describes an open deformable-asset schema developed with AOUSD and experimental coupling between solvers. In Isaac Lab benchmarks, he reports 2 to 3 times more simulation steps per second and 3 to 4 times lower memory use than Isaac 2 with PhysX.

4. 25:27

   Lightwheel’s assets and evaluation workflow

   Louis says Lightwheel builds and calibrates rigid and deformable assets, tests demanding simulations, and uses Newton in its Robot Finals evaluation product. Examples include a Samsung water-hose assembly, industrial cable insertion, and a cable whose simulated vibration and fall closely matched real footage. The team combines calibrated assets, robot system identification, teleoperation data, and parallel model evaluations, then uses reports to inform training.

5. 38:07

   Roadmap and invitation to contribute

   Miles summarizes Newton as an open-source engine with reusable sensors, collision tools, and pluggable solvers, and says the project has more than 100 direct contributors. He says Newton 1.4 includes deformable and coupling improvements, with Isaac Lab 3 availability planned for August. Future work mentioned includes asset migration tools, solver schemas, calibration tools, determinism, differentiability, and additional solvers.