Research
Notes from the lab.
Long-form thinking on simulation, perception, and policy learning for dexterous robots.
·Simulation
13 min read
Closing the contact gap with learned soft-body priors
Why ray-traced cloth and tendon simulations were the wrong abstraction, and how a diffusion video prior gets us photoreal manipulation without a render farm.
Read →·Perception
11 min read
MANO at scale: from monocular video to dense hand pose
A study of label noise, hand–object occlusion, and how 3D keypoint quality compounds into policy reward downstream.
Read →·Policy
15 min read
VLAs do not need more parameters — they need contact
Fine-tuning Pi-0 and OpenVLA on Abundance episodes shows that force-aware demonstrations beat 10× more vision-only data on real-world success rate.
Read →·Engineering
10 min read
LeRobot v2.0, why we ship it natively, and what we extended
An end-to-end walk-through of our exporter, including the contact-force channel and the 6-DoF object-pose registry.
Read →