UW and NVIDIA's D&A Trains Robots from 10% to 100% Success
D&A turns any VLM into a zero-shot robot reward model via a two-prompt decomposition. Paper A Describer narrates the video blind to the task; an Assessor scores the description against an explicit rubric.
Takeaways
- D&A turns any VLM into a zero-shot robot reward model via a two-prompt decomposition. Paper
- A Describer narrates the video blind to the task; an Assessor scores the description against an explicit rubric.
- Fixes confirmation bias where VLMs hallucinate progress on failed rollouts, yielding clear success vs failure separation.
- Lifted a DROID Franka picking policy from 10% to 100% success in ~90 minutes of online RL.
- Tops D&A-Bench, RBM-EVAL-OOD, and RoboRewardBenchmark, matching or beating post-trained reward models.
- Code and dataset released: GitHub , D&A-Bench .
A University of Washington and NVIDIA paper introduces D&A paper , short for Describe & Assess, a prompt-only method for scoring robot behavior with vision-language models (VLMs). D&A makes two sequential model calls: one describes the video without seeing the task, and the other compares that description with the instruction. The separation reduces false reports of progress when a robot drops an object, chooses the wrong item, or otherwise fails.
Robot-learning pipelines use reward signals to rank demonstrations and improve policies through reinforcement learning (RL). A dense reward assigns a progress score throughout a trajectory, giving the learner more information than a final success label. Hallucinated progress can promote failed trajectories, so reward accuracy directly affects what the robot learns.
Instructions bend the score
Direct reward prompts show a VLM the instruction and video together, encouraging the model to search for evidence that the requested task occurred. Existing zero-shot methods such as GVL and TOPReward can produce similar progress curves for successful and failed attempts. They may also behave like binary success detectors, missing partial progress and recovery.
Unreliable progress estimates can corrupt RL updates and behavior-cloning filters. Behavior cloning trains a policy to imitate recorded actions, often after selecting the strongest demonstrations. Post-trained systems such as Robometer, RoboReward, and SOLE-R1 address reward quality with specialized training, but they require robot data and additional training for each underlying model.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
