Humanoid robotics.

What

Use-case programming on Unitree H1-2 and G1 humanoid platforms for PepsiCo: defined industrial tasks deployed onto legged hardware, with locomotion governed by learned neural policies rather than scripted gait.

How

Policies were trained for legged control in simulation, then carried across the sim-to-real gap: system identification on the actuators, characterising where physical torque response and compliance diverge from the simulated model, and instrumenting the failure modes that only present on hardware. A controller that behaves in simulation comes apart against a real actuator, so the actuator is identified before the controller is trusted.

Results

Use cases delivered on both platforms, and a transfer methodology, identify the system before trusting it, that has carried into every system built since.