Artificial Computation for Bipedal Locomotion: Tokenized Walking Control and Semantic Stability

We present a theoretical framework for \textit{Artificial Computation} (AC) applied to bipedal locomotion, emphasizing high-level tokenized decision-making and semantic evaluation of walking stability. Unlike traditional low-level physics-based controllers, AC externalizes walking coherence into discrete action tokens, allowing adaptability across morphologies and environments. We formalize the AC paradigm, define the state space, attractor dynamics, parametric action tokens, and semantic reward functions, and illustrate a comparison to reinforcement learning-based approaches, highlighting robustness and modularity advantages.