Platform
The data engine for dexterous manipulation.
Four pieces, one pipeline — from a sentence to a fine-tunable dataset.
01 / Data engine
Prompts to physics-grounded episodes.
A constrained LLM samples scenario matrices across lighting, contact, and pose. Trajectories are validated in MuJoCo before any pixel is rendered, so every episode is physically plausible by construction.
02 / Simulation
Diffusion video, not ray-traced shortcuts.
Photoreal egocentric RGB-D produced by a learned video prior conditioned on simulated geometry. Cuts the sim-to-real gap without burning thousands of GPU-hours per minute of footage.
03 / Models
Drop into any VLA pipeline.
Exports as LeRobot v2.0 — joint angles, MANO keypoints, 6-DoF object pose, contact forces, depth. Compatible with OpenVLA, Octo, Pi-0, and in-house diffusion policies.
04 / Studio
Configure, generate, and ship.
A control surface for researchers and ML engineers. Define scenarios, queue batches, inspect failures with per-frame overlays, and download a versioned dataset when you are ready.
005 / The invitation
Build with abundance.
Early access is opening for research labs and frontier robotics teams. Tell us what you're training — we'll point you at the data.