Robot Dog Learns to FEEL the Ground — No RL, Biological Brain | MH-FLOCKE скачать в хорошем качестве

Robot Dog Learns to FEEL the Ground — No RL, Biological Brain | MH-FLOCKE

A Unitree Go2 robot dog learns to walk on hilly terrain using a biological brain with foot contact sensors — no reinforcement learning. NEW IN THIS VERSION: Foot contact sensing — the robot feels the ground with 4 feet Terrain reflexes — automatic slope compensation, like a vestibulospinal reflex Enhanced cerebellum — 2.3x stronger motor corrections Adaptive gait — slows down on slopes, pushes harder uphill The brain uses: Central Pattern Generators (innate gait, like a spinal cord) 5000+ spiking neurons (Izhikevich model, R-STDP learning) Cerebellar forward model (Marr-Albus-Ito, predicts motor consequences) Foot contact sensors (per-foot force + timing detection) Terrain reflexes (pitch/roll → immediate leg adjustment) Motivational drives + olfactory navigation All in a 15-step closed loop running at 200 Hz Results at 50k steps on hilly terrain: Max distance from start: 23.8m (reached at step 35k) The robot then navigates back toward start via olfactory steering Falls: 0 Actor competence: 1.0 (SNN fully in control, CPG only 40%) The Adaptive Terrain Response (ATR) system has three biological timescales: 1. Reflexes (ms): foot hits ground → immediate leg adjustment 2. Cerebellum (seconds): forward model learns terrain pattern 3. Behavior (minutes): gait adapts to slope steepness Next steps: Recovery learning (self-righting after falls) and stair climbing. Paper: "MH-FLOCKE: Biologically Grounded Embodied Cognition" — preprint on aiXiv Project (coming soon): https://www.mhflocke.com Contact: info@mhflocke.com Named after Flocke 🐕 Chapters: 0:00 Title Card 0:03 Motor Babbling — First Movements 0:15 Walking Begins — CPG Active 0:28 5m Reached — SNN Taking Over 0:38 Trotting — Terrain Reflexes Active 0:50 10m Reached — Cerebellum Adapting 1:02 SNN Takes Over — CPG Below 50% 1:10 20m — Full Terrain Adaptation 1:18 Final Trot — Actor Competence 1.0 1:22 End Card — Results #BiologicalAI #UnitreeGo2 #Robotics