A Biomimetic Dual-Brain Architecture for Robotics: Bridging Large Language Models and Reactive Control through Control Barrier Functions, Experience Memory, and Entropy-Guided Fine-Tuning

Large language models offer strong semantic reasoning, task decomposition, and zero-shot generalization for embodied AI, but their outputs remain probabilistic text or symbolic sequences, whereas robot control depends on millisecond-level state feedback, constraint enforcement, and execution updates. This gap between semantic planning and physical execution makes direct action generation vulnerable to hallucinated commands, invalid skill invocation, and misjudgment of embodiment-specific capabilities, while low-level reactive control alone is insufficient for long-horizon behavior in open environments. We address this problem with a dual-brain architecture composed of a contract-constrained planning layer, a local safety filter, an episodic experience buffer, and a hardware-aware adaptation supervisor. Skill primitives, units, and parameter bounds are derived from the robot’s local real-time control stack and define the contract space available to the upper-level model. The local safety filter, implemented near the robot body on an MCU, RTOS, or comparable embedded controller, applies control barrier functions (CBFs) and quadratic programming (QP) to keep executed commands inside a modeled safe set. The episodic experience buffer maintains both a knowledge index for rules, manuals, maps, and task templates and an embodied memory store for execution traces, intervention signals, and outcome feedback. The adaptation supervisor extracts low-intervention valid contracts from execution logs and performs hardware-aware parameter-efficient fine-tuning through LoRA- or QLoRA-style adaptation with entropy-guided weighting. The main contributions of this paper are a verifiable dual-brain systems model, a contract-based interface between natural language and safety-critical control, an intervention metric reusable by both retrieval and adaptation, and a reproducible evaluation protocol for safety, contract compliance, intervention decay, and cross-platform transfer.