SPEAR: Selective Preference Elicitation with Adaptive Rewards for VLM-Guided Robotic Manipulation
SPEAR transforms Vision-Language Models into efficient preference oracles for robotic manipulation by solving two main problems: expensive queries and poor initial exploration. It uses distinguishability-aware query selection to target only informative comparisons, reducing abstentions. It also employs an adaptive reward transition that uses VLM guidance for early exploration, allowing VLMs to be used practically without manual reward engineering.