Proceedings of the Annual Meeting of the Cognitive Science Society

This study explores how agents achieve stable locomotion in complex terrain. They do this relying only on proprioception, without visual guidance. We propose the Asymmetric Reinforcement Learning Contrastive Optimization (ARLC) framework. This framework simulates predictive coding in biological neural systems through a contrastive learning mechanism. The core of this model is to establish temporal consistency between historical action sequences and future expected actions. Experimental results show that this model has excellent generalization ability in the physical world. This includes zero-shot transfer from simulation to reality. More importantly, its learned latent representation space shows a self-organizing mapping of terrain categories. This provides a new computational perspective for understanding the evolution of internal representations in biological motor control.