Discrete Actions for Naturalistic Behavior in Embodied Biomechanical Animal Models

Embodied symbolic systems depend on a discrete vocabulary that compresses continuous motion into named actions, which make planning and annotation tractable. Many such vocabularies exist, from unsupervised segmentation to hand annotation, but they only describe motion, and making a body perform one of their symbols requires a hand-engineered controller or a policy trained for that symbol. We present Code2Act, which turns any discrete vocabulary into motor commands for a 38 actuator biomechanical rat. One recurrent controller reads only the symbol stream and proprioception, predicts the motor latent that a frozen pretrained imitation model would supply, and decodes through that model's inverse-dynamics module. A single training run then covers an entire vocabulary with no per-symbol reward and no per-symbol training. We drove one controller from three vocabularies of different provenance and size. An independent labeler can recover each steering symbol from the motion it produced, across arbitrary initial poses and external velocity impulses. Symbols are executed by the body and realized as distinct motion, and novel combinations of symbol pairs can produce within-distribution motion patterns.