From Vulnerability to General Intelligence: Autonomous Emergence of Latent Manifolds, Motor Primitives, and Active Inference under 3D Physical Dynamics

Evaluating the emergence of adaptive internal representations and autonomous behavioral regulation under existential physical constraints is a foundational problem in embodied artificial intelligence and cognitive science. Standard machine learning paradigms typically evaluate optimization under unconstrained, disembodied loss functions or episodic resets where failure incurs zero permanent penalty. In this work, grounded in Life-Information Theory, we demonstrate that non-resettable physical elimination (Darwinian mortality, HP <= 0), metabolic thermodynamic scarcity, and continuous nociceptive feedback are sufficient inductive constraints to drive the autonomous emergence of homeostatic latent manifolds, motor primitives, and communicative coordination in recurrent neural agents. We synthesize established 2D planar findings with a fully articulated 3D continuous physical simulation governed by Newtonian mechanics, gravity, contact friction, 16-sector spherical raycasting, and GPU-driven foveal saccadic vision. Key Theoretical & Empirical Contributions: Coupling Active Inference and Local Hebbian Plasticity: We formalize a dual-process interface connecting high-level variational free energy minimization with low-level reward-modulated Hebbian eligibility traces, establishing real-time sensorimotor adaptation and policy selection without global backpropagation through time (BPTT). Offline Sleep Consolidation & Basal Ganglia Crystallization: We model the offline crystallization of basal ganglia macro-organs via hippocampal sharp-wave ripple (SWR) replay over 480 consolidation cycles, executing 68,914 long-term potentiations (LTP) and 68,641 long-term depressions (LTD). This compresses fine-grained joint kinematics into reusable macro-action primitives, operating within a compact memory footprint (1.2 MB online memory, expanding temporarily to 8.4 MB during replay). Standardized 3D Physical Benchmarking: In standardized 3D continuous physics benchmarks (including gravity, contact friction, and 3-DoF momentum), our hybrid architecture sustains an operating survival rate of 79.6 ± 0.8% with low collision trauma (4.1 ± 0.3%), compared to the catastrophic collapse of standard deep RL baselines (PPO: 14.2%, SAC: 19.5%, MADDPG: 26.5%) and static neuroevolution (31.2%). The entire agent operates within an ultra-lightweight 459K-parameter footprint at 0.24 ms/step inference latency. Hierarchical Motor Control & 3D Spatial Attractors: Systematic 4-tier ablations demonstrate that decoupling low-road deterministic reflex arcs from high-road descending curiosity leads to severe starvation (38.2% survival) or catastrophic impact collisions (24.6% survival). Under 3D surveillance constraints, mortal agents exhibit an 8.0 ± 0.6% rate of physically wedging a 3D movable rigid block to mechanically jam the emergency termination switch (0.0% in non-mortal controls, p = 0.003). 3D Directional Auditory Coordination: Equipped with 3D directional audition under partial observability, agents coordinate multi-agent alerting networks, elevating cooperative depot harvesting yield to +68.4 ± 1.1% (vs. +21.2 ± 0.8% in deaf controls, p < 10^-25) and modulating ground-level evasive maneuvers against aerial predators. Reproducibility: Includes an explicit Open Science and Reproducibility Protocol with deterministic pseudorandom seed control, exact physical parameters, and an independent replication checklist. (This work scales and unifies the embodied frameworks established in Yoshikai, 2026, DOI: 10.5281/zenodo.24136124 and DOI: 10.5281/zenodo.24136607).