Environments
Simulation environments for embodied AI.
We build world models and simulation infrastructure that turn real-world inputs into structured, interactive 3D worlds.
Build an environmentFrom capture to interaction
Build the world, then make it useful for learning.
A simulation environment needs more than a visually convincing mesh. It needs structure an agent can perceive, change and learn from. We preserve that structure from source input through to the final interactive world.
- 01Inputs
Images, video, scans, CAD and floor plans.
- 02Geometry
Reconstructed spaces and assets with useful scale, topology and collision.
- 03Appearance
Materials, textures and lighting that preserve the visual evidence in the source.
- 04Understanding
Objects, rooms, relationships, dimensions, semantics and affordances represented explicitly.
- 05Behaviour
Articulation, physics and interactions that let agents act inside the world.
Capabilities
Three layers of a simulation-ready world.
Simulation-ready assets
Turn product imagery, scans and reference media into clean 3D assets with geometry, materials, articulation and collision prepared for simulation.
Structured scenes and digital twins
Reconstruct environments as inspectable scenes where rooms, objects, relationships, dimensions and physical constraints remain explicit.
Task environments for embodied agents
Build interactive worlds for training and evaluating navigation, manipulation, planning and other long-horizon embodied behaviours.
Infrastructure
Worlds that can be replayed, varied and measured.
The environment is part of the evaluation system. Every world and run must be reproducible enough to compare, while remaining varied enough to test genuine generalisation.
Deterministic reset
Recreate the same initial world for fair comparison and reliable debugging.
Procedural variation
Change layouts, assets, lighting and task conditions without reducing the world to one memorised path.
Versioned worlds
Track the exact assets, scene graph, physics and task configuration behind every run.
Telemetry
Record observations, actions, collisions, state changes, cost and latency as one inspectable trace.
Evaluation hooks
Score outcomes and process against geometry, state, safety and task-specific constraints.
Simulator-ready exports
Deliver assets and worlds in the formats required by downstream robotics and simulation stacks.