D&A: Decoupling Description and Assessment Enables VLMs as Zero-Shot Robotic Reward Models

This research introduces D&A (Describe & Assess), a prompt-based method that decouples task progress estimation into an objective textual description of scene dynamics followed by an assessment phase using explicit rubrics. By requiring the Vision-Language Model to synthesize an absolute definition of success before scoring, the approach anchors evaluations to external criteria rather than internal video timelines. Empirical results indicate that D&A calibrates zero-shot estimates and mitigates confirmation bias, providing fine-grained partial progress metrics that penalize failures while interpolating successful runs. The method generalizes across tasks and embodiments, matching or surpassing post-trained reward models, and enables efficient downstream reinforcement learning.

Generated from attributed reports · Updated 2 hours ago

There is not enough continuous observation data to show a trend.