mechatronics & applied robot learning

Robotics lead at 120ft factory. Previously founding engineer at k-scale labs (YC W24): first ecosystem of open-source humanoid robots, 200+ units pre-ordered for $2M+.

— why?

machines should add more to the world than they take from it, people and nature both richer for it. until then, i keep building.

built the robotics stack for a fleet of humanoids learning assembly for small-batch, high-variety manufacturing. Teleop, sim, and policy rollouts with interventions all grow one dataset; the robots put in 6-hour days at 60–70% subtask success.

started as a skunkworks project with two other engineers over winter break, tearing down our old robots to understand why they failed. Four months later it was the company flagship: the first open-source humanoid to deploy RL walking and VLA manipulation, 35 units in the wild.

a 2-ft humanoid for learning RL control. Shipped the first batch fast, watched 70% fail at a hackathon, rebuilt from what we learned. It went on to teach RL at stanford’s CS234, drew a 20k waitlist, and launched at CES.

— other projects

— odds and ends

Fine-tuned V-JEPA 2-AC to fold clothes on an openarm in closed loop within 24 hours: 1st place, GOSIM paris
Chasing the sim2real gap from both ends: measured physical parameters and contact geometry on hardware, URDFs and training environments in sim
A synthetic data pipeline that parameterizes scenes and objects, so VLA fine-tuning scales without more teleop hours
Egocentric video → mocap → whole-body AMP policies
3 technical hackathons hosted, the biggest past 1,000 attendees

— say hi: linkedin · twitter · github · email