GPU-accelerated evolutionary framework for studying morphable soft-body robots that approach, combine, and locomote as merged entities using 400-particle spring-mass systems controlled by evolved neural networks. v2 updates (March 6, 2026): Experiment 5: Generalization across asymmetry types (stiffness, shape, combined). Establishes the Inertial Asymmetry Principle—only mass asymmetry drives differentiation; stiffness (r=0.579) and shape (r=0.735) fail to break synchronization. Experiment 6: Differentiation Dynamics over 1000 generations. Discovers the V-shaped transient differentiation trajectory: rapid specialization at ~25 generations (r=0.056), followed by partial re-synchronization into a stable coordination plateau (r≈0.58) as fitness improves by +146%. Self-evolving morphology: when masses are evolvable genes, evolution maintains symmetry (ratio 1.11:1), confirming differentiation requires external imposition. Key findings (5 principles): Symmetry Locks Theorem: Symmetric bodies cannot differentiate (r=0.742). Inertial Asymmetry Principle: Only mass asymmetry breaks synchronization (Mapping Inequivalence via F=ma). Transient Differentiation: Role specialization peaks early then is partially reabsorbed for efficiency. Reward Paradox: Multi-phase fitness can mask poor locomotion via reward hacking. Symmetry Optimality: Evolution chooses symmetric masses when free to evolve morphology. All experiments run on a single NVIDIA RTX 5080 Laptop GPU in under 1 hour total.
Hiroto Funasaki (Fri,) studied this question.