Hardlight Robot

Physics

Encoded

01

Synthesis

From a single pixel input, we infer volumetric depth. Our model generates clean, watertight 3D topology from any 2D reference image.

02

Physics

We automatically compute inertial properties. Mass, friction coefficients, and collision bounds are encoded based on inferred material composition.

03

Simulation

Zero manual rigging. Assets export as native USD or URDF, validated for stable contact dynamics in NVIDIA Isaac Sim and MuJoCo.

Built for Physically Accurate Training

Robotic Manipulation

Train agents to interact with complex articulation. We encode precise collision geometry for doors, drawers, and tools.

Navigation

Generate diverse, interactive clutter. Create chaos-ready environments for mobile robots to master pathfinding.

Domain Randomization

Prevent overfitting with infinite variations. Randomize textures, lighting, and mass properties instantly.