MH-FLOCKE is a biologically grounded embodied AI system controlling a Unitree Go2 quadruped robot in MuJoCo simulation. Every component has a biological equivalent — from spiking neurons to cerebellar forward models.
Architecture OverviewThe 15-step cognitive cycle from sensors to motor output.
SNN Controller4,500 Izhikevich neurons with neuromodulation.
R-STDP LearningFree Energy Principle meets reward-modulated plasticity.
Cerebellar LearningMarr-Albus-Ito forward model with multi-compartment PkC.
Spinal CPGPhase-coupled oscillators with competence gate.
Task PE & Ball InteractionLoss aversion and DishBrain-style stimulation.
Reflexes & TerrainSpinal segments, righting reflex, terrain adaptation, VOR.
Emotions, Drives & BehaviorSomatic markers, four drives, behavior planner.
World Model & DreamsSpiking predictive model with offline replay.
Global Workspace (GWT)Attention competition between cognitive modules.
Body SchemaEfference copy and forward model for self-awareness.
Episodic MemoryEmotionally indexed trajectory fragments with recall.
MetacognitionSelf-monitoring, consciousness level (0–12).
Curiosity & EmpowermentIntrinsic motivation via prediction error and action influence.
Synaptogenesis & ConceptsSNN patterns consolidated into a knowledge graph.
Consistency & IntegrityDissonance detection with neuromodulator reset.
Developmental ScheduleMotor maturation from babbling to skilled locomotion.
Population CodingGaussian tuning curves and push/pull motor decoding.
Physics & TerrainMuJoCo integration, heightfields, ball injection.
Scene & Task ParsingNatural language to training configuration.
Brain PersistenceSave and load complete cognitive state.
Purkinje CompartmentsMulti-compartment dendritic computation with calcium.
Skills & Meta-LearningEWC skill protection and evolved plasticity rules.
Training PipelineCLI, reward, curriculum, checkpointing.
FLOG Format & DashboardBinary log format, analysis server, live monitoring.
SNN Controller4,500 Izhikevich neurons with neuromodulation.
R-STDP LearningFree Energy Principle meets reward-modulated plasticity.
Cerebellar LearningMarr-Albus-Ito forward model with multi-compartment PkC.
Spinal CPGPhase-coupled oscillators with competence gate.
Task PE & Ball InteractionLoss aversion and DishBrain-style stimulation.
Reflexes & TerrainSpinal segments, righting reflex, terrain adaptation, VOR.
Emotions, Drives & BehaviorSomatic markers, four drives, behavior planner.
World Model & DreamsSpiking predictive model with offline replay.
Global Workspace (GWT)Attention competition between cognitive modules.
Body SchemaEfference copy and forward model for self-awareness.
Episodic MemoryEmotionally indexed trajectory fragments with recall.
MetacognitionSelf-monitoring, consciousness level (0–12).
Curiosity & EmpowermentIntrinsic motivation via prediction error and action influence.
Synaptogenesis & ConceptsSNN patterns consolidated into a knowledge graph.
Consistency & IntegrityDissonance detection with neuromodulator reset.
Developmental ScheduleMotor maturation from babbling to skilled locomotion.
Population CodingGaussian tuning curves and push/pull motor decoding.
Physics & TerrainMuJoCo integration, heightfields, ball injection.
Scene & Task ParsingNatural language to training configuration.
Brain PersistenceSave and load complete cognitive state.
Purkinje CompartmentsMulti-compartment dendritic computation with calcium.
Skills & Meta-LearningEWC skill protection and evolved plasticity rules.
Training PipelineCLI, reward, curriculum, checkpointing.
FLOG Format & DashboardBinary log format, analysis server, live monitoring.