Driving · 2026-11-03
D&A: Decoupling Description and Assessment Enables VLMs as Zero-Shot Robotic Reward Models
Marius Memmel, Shawn Yin, Urvi Rutia, Angad Wadhwa, Eldon Li, Gavin Angeloff, Ankit Goyal, Jesse Zhang, Dieter Fox, Fabio Ramos, Anqi Li, Abhishek Gupta
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors propose D&A (Describe & Assess), a purely prompt-based zero-shot reward model for robotic video task-progress estimation. It decouples a VLM-based trajectory description (without the task) from an assessment stage that uses an explicit, text-based success criterion/rubric to score per-frame progress in [0,1]. This mitigates confirmation bias and yields dense partial progress rewards and rich failure feedback for downstream RL.
Why it matters
The paper targets confirmation-bias failures of naïve zero-shot VLM prompting, where models often hallucinate progress on failed executions. By anchoring scoring to an explicit global success criterion and separating description from assessment, D&A aims to produce calibrated, fine-grained rewards useful for policy learning and failure-aware feedback, while remaining zero-shot and model-agnostic.
Method
- Decouple reward prediction into a describer that narrates robot behavior from the video unconditioned on the task instruction, then an assessor that scores the description against a success criterion generated from the instruction.
- Generate an explicit natural-language success criterion within the assessor, then score each timestep relative to it to enable partial progress and resist reward-collapse modes.
- Optionally augment the assessor with feedback generation to output failure analysis and corrective instructions alongside scalar progress rewards.
Abstract (from arXiv)
Vision-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. Naive prompting of frontier models collapses into a binary success detector that fails to capture nuanced, non-linear progress and failures. While post-training approaches can mitigate this through explicit training, they demand expensive real-world robotics data and require computationally intensive post-training for every individual base model. To resolve these limitations, we introduce D&A (Describe & Assess): our approach decouples progress estimation into an objective, task-grounded textual description of scene dynamics followed by an assessment phase using explicit rubrics. By compelling the VLM to synthesize an explicit, text-based definition of absolute success before scoring, the framework anchors evaluations to an external, global criterion rather than the video's internal timeline. Empirical results demonstrate that our method calibrates zero-shot estimates and mitigates confirmation bias, providing fine-grained, partial progress metrics that accurately penalize failed executions while correctly interpolating successful runs. Being purely prompt-based, D&A generalizes across tasks, embodiments, and benchmarks---matching or surpassing post-trained reward models---and its reward estimates enable efficient downstream reinforcement learning.
Related papers
- RoboSeed: Multidimensional Evaluation of Vision-Language-Action and World-Action Models: Capability, Resource Affordability, and Research Evidence
- Predictive Long-Horizon Progress Modeling with Action-Aware Quantification
- Video2STL: Grounding VLM-Generated Temporal Specifications for Robot Learning
- Learning Robotic World Models with Semantic and Visual Feedback
- Causal Analysis of Robot Foundation Models Enables Efficiency Gains
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-11-02 · Indexed by Pub-AI: 2026-11-03 UTC
D&A: Decoupling Description and Assessment Enables VLMs as Zero-Shot Robotic Reward Models
Preprint posted: 2026-11-02 · Original record ↗
- arxiv
- 2611.03622
Explicit source versions (1)
- v1 · 2026-11-02
- Preprint available
Source date: 2026-11-02 · First detected: 2026-11-03 UTC · Historical baseline
Source evidence (2)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2611.03622v1; arxiv:announce_type=new; item pubDate (official metadata)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Vision-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. Naive prompting of frontier models collapses into a binary success detector that fails to capture nuanced, non-linear progress and failures. While post-training approaches can mitigate this through explicit training, they demand expensive real-world robotics data and require computationally intensive post-training for every individual base model. To resolve these limitations, we introduce D&A (Describe & Assess): our approach decouples progress estimation into an objective, task-grounded textual description of scene dynamics followed by an assessment phase using explicit rubrics. By compelling the VLM to synthesize an explicit, text-based definition of absolute success before scoring, the framework anchors evaluations to an external, global criterion rather than the video's internal timeline. Empirical results demonstrate that our method calibrates zero-shot estimates and mitigates confirmation bias, providing fine-grained, partial progress metrics that accurately penalize failed executions while correctly interpolating successful runs. Being purely prompt-based, D&A generalizes across tasks, embodiments, and benchmarks---matching or surpassing post-trained reward models---and its reward estimates enable efficient downstream reinforcement learning.
