D&A Enables VLMs as Zero-Shot Robotic Reward Models
Original title: D&A: Decoupling Description and Assessment Enables VLMs as Zero-Shot Robotic Reward Models
Decouples progress estimation into objective description and rubric-based assessment to reduce confirmation bias in zero-shot VLM evaluations for robotics.
Vision-Language Models promise zero-shot video task-progress estimation in robot learning yet suffer from confirmation bias and hallucinating progress. The D&A method decouples the task into an objective textual description of scene dynamics followed by explicit rubric assessment. This anchors evaluations to external criteria and provides fine-grained partial progress metrics without post-training or real-world data.
- Mitigates confirmation bias by anchoring to global criteria
- Delivers nuanced non-linear progress metrics
- Generalizes across tasks, embodiments and benchmarks
- Enables efficient downstream RL
- Matches or surpasses post-trained reward models
From the original article
1 min read at arxiv.orgVision-Language Models (VLMs) hold promise for zero-shot video task-progress estimation in robot learning, yet existing frameworks frequently suffer from confirmation bias, hallucinating progress in failed executions due to implicit success assumptions or relative temporal ranking designs.
By Marius Memmel · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.