D&A: Decoupling Description and Assessment Enables VLMs as Zero-Shot Robotic Reward Models
D&A scores partial robot-task progress and penalizes failed attempts using prompts alone, the authors report.
The method first produces a task-grounded description of scene dynamics, then assesses progress against explicit rubrics. Before scoring, it defines absolute success in text, anchoring judgments to that criterion rather than the video’s timeline. The authors report reduced confirmation bias, including less imagined progress in failed executions. Across their evaluations, D&A matches or surpasses post-trained reward models, and its scores support downstream reinforcement learning.
ArXiv · AI/CL/LG's note
The method first produces a task-grounded description of scene dynamics, then assesses progress against explicit rubrics. Before scoring, it defines absolute success in text, anchoring judgments to that criterion rather than the video’s timeline. The authors report reduced confirmation bias, including less imagined progress in failed executions. Across their evaluations, D&A matches or surpasses post-trained reward models, and its scores support downstream reinforcement learning.
ArXiv · AI/CL/LG's note
score 5