QuickLAP: Quick Language-Action Preference Learning for Semi-Autonomous Agents

In three linesQuickLAP fuses physical and language feedback to learn robot reward functions in real time using a Bayesian framework. LLMs extract reward feature attention masks and preference shifts from free-form utterances, integrated with physical corrections via closed-form update rule. Achieves 70% error reduction vs physical-only and heuristic multimodal baselines in semi-autonomous driving simulator.

Summary generated by Claude — human-verified