RareSense/Nova3D

What it solves

Nova3D creates high-fidelity, physics-accurate 3D environments for robotics simulation. Unlike traditional 3D mesh generators that produce "vertex soup" (static, fused meshes), Nova3D generates worlds as executable scene programs. This ensures that objects have real hierarchies, working mechanisms (like hinged doors or sliding drawers), and precise physical properties (mass, friction, and inertia) required for training and testing robots in simulation.

How it works

The system converts text prompts into executable construction procedures. These programs are then compiled into formats compatible with major robotics simulators such as MuJoCo, Isaac Sim, Gazebo, and PyBullet.

Key technical features include:

• Mechanism Validation: Every generated mechanism (hinge, slider, etc.) is force-tested by the physics engine to ensure it doesn't jam or float before the scene is shipped.
• Digital Cousins: The ability to rebuild real-world locations from photographs, ensuring every object is sized and placed accurately against the original.
• Multi-Engine Export: A single source of truth for the environment is exported to multiple native formats (MJCF, USD, SDF, URDF) to maintain consistency across different simulation platforms.

Who it’s for

Robotics researchers and developers who need realistic, verifiable physics environments for training AI agents, testing walking policies, or validating robot interactions with the world.

Highlights

• Physics-First Generation: Generates environments with native gravity, contact, and friction properties rather than just visual approximations.
• Working Mechanisms: Produces separately addressable bodies and joints that robots can actually interact with (e.g., opening a door).
• Verified Builds: Uses automated force-tests and terrain probe-tests to gate every build.
• Photo-Matched Captures: Creates "digital cousins" of real places for high-fidelity simulation.
• Broad Compatibility: Native exports for MuJoCo, Isaac Sim, Gazebo, and PyBullet.