MH-FLOCKE MH-FLOCKE
HomeDocsGitHubBlogPapersYouTubeReddit𝕏

Documentation

MH-FLOCKE is a biologically grounded spiking neural network system for quadruped robot locomotion. No deep reinforcement learning, no gradient descent — the robot learns from embodied experience using the same principles biological nervous systems use.

The current public release targets the Petoi Bittle X (535 neurons) — an 8-DOF quadruped, in MuJoCo simulation and on real hardware over a WiFi bridge, with the SNN learning on-device via R-STDP. The same architecture and code also run on a Unitree Go2 in simulation (1,376 neurons) and a Freenove Robot Dog on a Raspberry Pi 4 (560 neurons, ~5 watts) — the platforms documented in Papers 1 and 2. Same architecture, same code, different bodies.

Key Results

Config Distance (m) Falls Variance
SNN + Cerebellum 45.15 ± 0.67 0 σ = 0.67
CPG only 40.73 ± 6.14 0.2 σ = 6.14
PPO Baseline 12.83 ± 7.78 0 σ = 7.78

3.5× further than from-scratch PPO with 11.6× lower variance. 10-seed validation on Unitree Go2, 50k steps each. Zero falls across all seeds.

How to read this honestly: the margin over PPO is mostly the innate CPG locomotion prior, not the SNN learning to walk by itself. The SNN + cerebellum’s own contribution over CPG-only is about +11% distance, alongside the variance collapse (σ 6.14 → 0.67) and zero falls. “From scratch” describes the PPO baseline, not these numbers.

What Makes It Different

  • Spiking neurons — Izhikevich model with biologically realistic dynamics, not differentiable activations
  • Cerebellar forward model — Marr-Albus-Ito architecture with a sparse granule expansion layer (Bittle 278) for motor prediction and correction
  • Central Pattern Generator — innate locomotion rhythm that the SNN learns to modulate, not replace
  • Reward-modulated STDP — learning through spike timing and dopamine, not backpropagation
  • PID closed-loop steering — IMU yaw feedback for drift compensation (Bittle --yaw-pid; asymmetric-stride phototaxis on Freenove)
  • Meta-learning loop — autonomous strategy adaptation through episode analysis, curiosity, and hypothesis testing
  • Runs on a Raspberry Pi — 38 steps/sec at ~5 watts, real-time on a 100€ robot kit

Documentation

  • Architecture — The 15-step cognitive cycle, from sensors to motor output
  • Hardware — Bittle WiFi bridge quickstart, plus Freenove (Paper 2) Pi deployment
  • Getting Started — Installation, first training run, rendering
  • Roadmap — What’s next for MH-FLOCKE
  • Changelog — Version history

Papers

Links