Data for the paper titled "A Schema Bounded Language Model for Refining Robot Policies Without Destabilizing Local Learning"

These processes operate at different timescales, yet a practical architecture must connect shared experience to local adaptation without transferring action selection to a central model. This work examines the decentralized reuse of policy experience among robots guided by heterogeneous language models and develops a hierarchical framework for repeated policy refinement while retaining robot-local learning and action selection.

Each robot owns a schema-bounded LLM policy agent, an Upper Confidence Bound (UCB) bandit, and a policy-conditioned Double Deep Q-Network (Double DQN) controller. At round boundaries, LLM agents exchange policy experience through a shared board, and UCB-guided refinement uses this context to update each robot’s local policy within a constrained schema. Within each round, Double DQN combines local state with the current language-derived policy to select actions through online correction.