Rainier Labs

Terrier

Terrier is a quadruped, and when I arrived it was a prototype that could not yet walk. I built its locomotion controller, working both sides of the problem at once: classical control for what could be derived, deep reinforcement learning for what could not.

Learning to walk

Gaits are hard to hand-write and easy to learn badly. I architected the training environments in NVIDIA Isaac Sim and wrote the Proximal Policy Optimization routine that ran in them, with 4000 Terriers spawned in parallel, all failing and correcting at once.

The routine converged within 12,000 episodes and lifted stable-gait success from 30% to 85%. I trained and evaluated policies across five terrain types including slopes, gravel and stairs, and the policies held 80% gait stability on surfaces they had never seen in training, which is the number that says the controller learned to walk rather than to memorize a floor.

Onto the hardware

A policy that works in simulation has proven nothing. I deployed the trained policy to the physical robot and validated it across more than 30 hardware-in-the-loop trials, where it held the same 85% stable-gait success it reached in simulation.

Off the tether

The last test is not a benchmark. It is a crowded room, a hard floor, and a real dog that has opinions about what a four-legged thing should move like.