Skip to content
← Projects

Spot Micro · Leika

A quadruped on an ESP32-CAM, with a learned policy correcting the hand-written gait.

Years
2023–2026
Status
Active
Field
hardware
GitHub
★ 79 · 21 forks

About

My take on the quadruped Spot Micro. An ESP32-CAM running FreeRTOS does inverse kinematics and gait generation on the device, and a SvelteKit app drives it over WebSocket with live telemetry and camera streaming. On top of that sits a residual walking policy trained in MuJoCo with Gymnasium and PPO: the firmware gait is ported to NumPy as the baseline, and the policy learns only small per-foot corrections. A zero action reproduces the hand-written gait exactly, so training starts from something that already walks and spends its capacity on staying upright.

Built with

  • Reinforcement learning
  • MuJoCo
  • C++
  • ESP32-CAM
  • FreeRTOS
  • SvelteKit