← Projects
Hexapod
A locomotion policy trained in simulation, running on the microcontroller itself.
- Years
- 2024–2026
- Status
- Active
- Field
- hardware
- GitHub
- ★ 6 · 2 forks
About
Eighteen servos on an ESP32-S3. Rather than hand-writing a gait, the walking controller is a neural network trained in a MuJoCo environment and then exported to a C header the firmware compiles in. It takes a 50-dimensional observation — body-frame gravity, gyro, joint angles, a gait phase clock and the previous action — and emits 18 joint targets at 50 Hz, on the device, with no host in the loop. Getting that to survive the crossing from simulation to real servos is most of the work.
Built with
- Reinforcement learning
- MuJoCo
- PyTorch
- C++
- ESP32-S3
- SvelteKit