A robot dog that learns to walk and navigate — not through programming, but through spiking neurons, a cerebellar forward model, and embodied experience.
The Petoi Bittle X is MH-FLOCKE’s current platform — shown here in MuJoCo simulation. The same brain runs on the real robot over a WiFi bridge and keeps learning on-device; on-hardware footage is in progress.
A brain trained in simulation transferred directly to the Freenove Robot Dog on a Raspberry Pi 4 — the sim-to-real result behind Paper 2.
MH-FLOCKE runs the same SNN and cerebellum code in the MuJoCo simulator and on real robots — one codebase, built from a shared topology profile. The current hardware platform is the Petoi Bittle X (~100€): the brain runs on a host PC connected over a WiFi bridge (~4 Hz), while the Bittle streams IMU data and executes joint commands.
On the robot the SNN keeps learning through R-STDP, driven by intrinsic reward — vestibular comfort, curiosity, empowerment — with no external reward and no pretraining required.
The approach was first proven end-to-end on the Freenove Robot Dog with a Raspberry Pi 4 (Paper 2): a brain trained in simulation transferred to the real robot via a single file and continued learning on the Pi at ~38 steps/sec on ~5 watts.
Steering closes the loop on the IMU. The Bittle holds a straight line with a yaw-PID controller; the Freenove navigates toward light via camera-based phototaxis, curving by varying stride length between legs — like tank steering. No GPS, no SLAM — an IMU, and for phototaxis a camera.
MH-FLOCKE is a scientific experiment: Can a robot learn to walk and navigate using biological principles instead of deep learning?
The same architecture runs across three quadrupeds: the Petoi Bittle X (535 neurons, the current hardware platform), the Freenove Robot Dog (560 neurons, the Paper 2 sim-to-real result), and a Unitree Go2 in MuJoCo simulation (1,376 neurons, the Paper 1 ablation). Same code, different bodies.
MH-FLOCKE implements spiking neural networks with R-STDP, a Marr-Albus-Ito cerebellar forward model, central pattern generators, and a Free Energy framework — all running simultaneously in a 15-step cognitive cycle.
The result: a quadruped that walks on an innate CPG gait and refines its balance and steering through experience, navigates toward light, and compensates mechanical drift — driven by spiking dynamics and intrinsic reward, without backpropagation and without a GPU.
A biologically grounded cognitive architecture — from spinal reflexes to meta-learning.
535–1,376 Izhikevich neurons in cerebellar populations. Granule cells, Golgi, Purkinje, DCN, Motor Hidden. Learning via R-STDP.
Marr-Albus-Ito architecture with a sparse granule expansion layer (278–567 cells). Climbing-fiber error signals drive LTD/LTP in Purkinje cells.
Central pattern generators produce rhythmic gaits. A competence gate blends the innate CPG with the learned actor (90% → 40% CPG).
IMU-based closed-loop drift compensation. On the Bittle a yaw-PID holds a straight line; on the Freenove a camera drives phototaxis via asymmetric stride.
Dopamine, serotonin, norepinephrine, acetylcholine dynamically modulate learning, exploration, and arousal.
Episode analysis, strategy adaptation, curiosity-driven exploration, and hypothesis testing. The dog improves its own learning strategy.
Go2 10-seed validation. The full stack walks 3.5× further than PPO — most of that distance from the innate CPG, with the SNN+cerebellum delivering 11.6× lower variance and zero falls. aiXiv 260301.000002
Freenove hardware deployment. Bridge v4.4, unified codebase, brain persistence across sessions. aiXiv 260409.000002