A platform for investigating prompt framing as interface parameters in foundation models for robotics

Foundation models, in particular large language models (LLMs), are finding increasing popularity when used in describing goals for robotic control, decision making, and execution. Recently, proposals for hybrid paradigms leveraging strengths of reinforcement learning (RL) agents in tandem with LLMs for robotic control have been demonstrated. The interface between the RL agents and the language model however offers a unique opportunity to explore how prompt framing may affect such hybrid systems. This work presents a controlled experimental platform to measure and better understand how manipulation of the interface between RL agents and an LLM impacts behaviour of a hybrid advisor-arbiter architecture. We compared three agents under matched evaluation protocols and initializations in a simulated navigation environment: (i) RL-only tabular Q-learning; (ii) LLM-only (stateless) action selection; and (iii) a hybrid LLM + RL agent. Under a constrained interaction budget (10 episodes per world), the hybrid LLM + RL agent achieves higher mean success and higher mean cumulative reward than both RL-only and LLM-only baselines. Advisor-channel ablations (random recommendations and null recom