From Vulnerability to Coordination: Autonomous Emergence of Homeostatic Latent Dynamics, Acoustic Signaling, and Extinction Resistance in Embodied Recurrent Agents
Year: 2026
Type: preprint
Abstract: Evaluating the emergence of adaptive internal representations under existential constraints is a foundational question in embodied artificial intelligence and cognitive systems. In this work, we investigate how non-resettable physical elimination (Darwinian mortality) and nociceptive feedback shape policy dynamics in low-dimensional recurrent neural agents (Gated Recurrent Units, H = 32) operating within a 2D continuous kinematic arena. Without pre-programmed emotional modules or symbolic communication seeds, agents exposed to somatic damage feedback autonomously structure latent manifolds satisfying operational criteria for threat-contingent coordination: cross-hazard representation convergence (cosine similarity 0.469 ± 0.06), anticipatory threat decoding under sampling truncation (ROC-AUC 0.642), and motor perturbation responses under latent vector injection (z = 0.625). Key Findings & Benchmarks: Empirical RL Benchmarking: Systematic comparison against standard reinforcement learning baselines (PPO, SAC, MADDPG, and Static Neuroevolution; N = 50 per condition) demonstrates statistically significant advantages in terminal survival (84.3% vs. 21.4–32.1%, p < 10^-80) and latency (0.12 vs. 4.9–15.4 ms/step) under non-resettable mortality. Acoustic Signaling & Channel Ablation: Under cooperative foraging with asymmetric sensory constraints, agents discretize continuous acoustic channels into distinct signaling centroids (MI = 0.122) functioning as stochastic alerting and recruitment beacons. Systematic channel ablations confirm that the +78.4% harvest gain decomposes into rear-predator alerting (+31.2%) and nutrient clustering (+42.6%). Spatial Dilemma & Ontogenetic Avoidance: In a spatial social dilemma, mortal agents exhibit a 97.0% individual retreat rate under elevated recurrent activation norms, reflecting the mathematical predominance of real-time Hebbian pain-avoidance gradients over unrepresented group-level payoffs. Conditioned Aversive States & Synaptic Dampening: Conditioned aversive associations exhibit persistent extinction resistance arising from the asymmetric reinforcement structure of Hebbian eligibility traces. Post-retrieval synaptic weight dampening acts as a phenomenological operator of attractor destabilization while preserving general motor competence (92.4% baseline foraging retention). Reproducibility: The work provides an explicit Open Science and Reproducibility Protocol with an independent replication checklist, detailing deterministic replication seeds and architecture parameters.
Author: Reiji, Yoshikai
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225825330
Cites:
Cited by:
Related to: