MH-FLOCKE: Biologically Grounded Embodied Cognition Through a 15-Step Closed-Loop Architecture for Quadruped Locomotion Learning

Abstract

I present MH-FLOCKE, an embodied AI platform in which simulated quadruped creatures learn locomotion through a biologically grounded cognitive architecture. Unlike end-to-end reinforcement learning (RL) approaches that treat the body as an optimization target, MH-FLOCKE implements a 15-step closed-loop processing cycle that integrates proprioception, embodied emotions, episodic memory, motivational drives, a Global Workspace for attentional competition, metacognitive self-assessment, and reward-modulated spike-timing-dependent plasticity (R-STDP) in a spiking neural network (SNN). In this revision, I address reviewer feedback by providing: (1) mathematical formulations of all core learning rules (Izhikevich dynamics, R-STDP, cerebellar forward model, competence gate), (2) a PPO baseline comparison showing MH-FLOCKE's neural learning core achieves 3.5× the walking distance (45.15 ± 0.67 m vs. 12.83 ± 7.78 m at 50k steps), (3) multi-seed statistical validation across 10 seeds and 80 runs, and (4) cross-embodiment transfer to the Unitree Go2 without architectural changes. Systematic ablation across 60+ runs isolates component contributions: vestibular reflexes eliminate all falls, motor babbling increases flat-terrain distance by 763%, the cerebellar forward model produces measurable prediction errors, and an olfactory sensory environment enables stimulus-driven behavior switching. The SNN+Cerebellum core (B1) achieves s = 0.67 m across 10 seeds — the lowest variance of any condition. I also report an interaction effect where the full cognitive architecture reduces locomotion distance compared to the neural core alone.

Keywords