ODIE V3: "Man's Best Friend"
By ODIE v3, I had the basics down from v1 and v2. Now I wanted to optimize. I trained reinforcement learning policies in MuJoCo simulation to discover the perfect gait, then deployed that policy directly to the real robot. Upgraded to a Raspberry Pi 5 for better compute, added a PCA9685 servo driver for smoother control, and used a Bambu Lab P1S 3D printer to fabricate better parts that reduced weight and friction. The hardware finally matched the software I wanted to run.
The RL policy learned gaits I never would have hand-coded. Trot diagonal patterns that were smooth, efficient, and actually graceful. Sim-to-real transfer meant the learned behavior worked on physical hardware without manual tuning or adjustment. This was the culmination of everything—raw iteration, simulation validation, and finally learned optimization. A quadruped that actually moves like it knows what it's doing.