Goal-conditioned quadruped navigation with a Mixture-of-Experts policy

• ›A custom Gymnasium environment around the ANYbotics ANYmal C model from the MuJoCo Menagerie. Observations are joint positions and velocities, actions are the twelve actuator torques, and an episode ends when the body drops below 20 cm.
• ›Stage one: locomotion. A reward of forward velocity plus a posture bonus minus an action-cost penalty, trained with PPO from Stable-Baselines3, with video captured every ten thousand steps so the gait can be watched improving from falling over to an early walk.
• ›Stage two: goal-conditioned navigation. The observation gains a target, and the policy becomes a Mixture-of-Experts network in PyTorch, with a gating network choosing among specialist experts per step. It reaches 92% success on target-reaching tasks within 200 timesteps.

Agents that act in the physical world have all the problems of software agents, with worse failure modes. This is the arm I keep pulling outside work.