Harnessing VLM in Embodied RL: From Reward Regressor to Success Detector via Exploration-Augmented Key-Observation Gating

This research introduces VLMeg, a method that utilizes Vision-Language Models (VLMs) as success detectors in embodied reinforcement learning. By aligning VLMs with their inherent multi-modal capabilities, the approach improves reward design and achieves state-of-the-art results in reward-free tasks. The method includes an exploration-augmented key-observation gating mechanism to enhance RL efficiency and detection accuracy.