001  /  Physical Intelligence

Synthetic Training Data
for Fine Motor Skills.

Describe a task. Receive physically-verified egocentric video, 3D hand poses, and contact forces — exported in LeRobot v2.0, ready to fine-tune any VLA.

001  /  Physical Intelligence

21 kp · MANO · 60 fps
LeRobot v2.0 native3D hand poseContact forcesEgocentric RGB-DPhysically verified240 fps captureMANO topology6-DoF object posePer-frame ground truthSim-to-real video priorLeRobot v2.0 native3D hand poseContact forcesEgocentric RGB-DPhysically verified240 fps captureMANO topology6-DoF object posePer-frame ground truthSim-to-real video prior

002 / The system

Three stages, one pipeline.

From a single sentence to a fine-tunable dataset — every step grounded in physics.

01

Describe

Write the task in plain language. A constrained LLM samples a structured variation matrix across lighting, placement, contact, and hand pose.

02

Simulate

MuJoCo plays the trajectory under physics. A diffusion video prior renders it as egocentric RGB-D with realistic motion blur and occlusion.

03

Export

Every frame ships with 21 MANO keypoints, 6-DoF object pose, contact forces, and per-pixel depth. Bundled as LeRobot v2.0 — paste it into your VLA.

003 / Outcomes

Numbers that move benchmarks.

0M

Episodes generated

0.0mm

Hand pose accuracy

<0h

Dataset turnaround

004 / Manifesto

The bottleneck for general-purpose robots was never compute it was data. The data we needed didn't exist, so we built the machine that makes it.

From the Abundance research notebook

Read the research →

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.