WATonomous / June to July 2026

GPU-Accelerated Humanoid Locomotion

I built the simulation and reinforcement-learning pipeline that taught WATonomous's custom humanoid lower body to walk before the physical legs were available.

LEGR humanoid walking in NVIDIA Isaac Sim
Isaac Sim / PPO rollout
Role
Primary developer, LEGR locomotion
Robot
Custom 12-DOF biped
System
Isaac Lab, RSL-RL, PPO
Validation
Fresh training from random initialization

The challenge

Building the bridge from CAD to learned control.

This was not an adaptation of a stock simulated robot. I owned the path from the team's SolidWorks-exported model to a trainable 12-DOF articulation in NVIDIA Isaac Lab.

I validated the kinematic chain, established symmetric joint behavior, configured joint-specific actuator dynamics and contact sensing, and created dedicated flat and rough terrain environments while preserving the mechanical model as the source of truth.

Final policy validation 19 seconds / Isaac Sim

Engineering approach

Designed as a system, not a single training run.

Embodiment

A robot the policy could trust.

I audited articulation, limits, mirrored joint response, root structure, contacts, and actuator behavior before training. This made the simulated mechanism a reliable interface for both the policy and the mechanical team.

Learning

Rewards grounded in observed failure modes.

Early policies stood still, shuffled, favored one leg, crossed their feet, rolled their ankles, and drifted off heading. I shaped velocity tracking, gait timing, swing-foot motion, bilateral spacing, balance, and control smoothness around those behaviors. Training ran across up to 256 GPU-parallel environments, with physics at 200 Hz and learned control at 50 Hz.

Diagnostics

Separating physics errors from policy errors.

I built focused tools for spawn stability, individual joint response, and policy rollouts. They exposed whether a failure came from articulation, contact physics, reward design, or learned behavior before I committed GPU time to another run.

Validation

The completed system relearned walking from a fresh initialization.

Checkpoint resumes made iteration faster, but they could also preserve poor local optima. I validated the final configuration with a new 10,000-iteration PPO run from random initialization.

The resulting flat-terrain policy balanced, alternated its legs, and produced repeatable forward locomotion in Isaac Sim. The implementation was authored by me, reviewed by the team, and merged as WATonomous PR #95.

Inspect the merged implementation