ProgressCompass: Embodied Progress Reward Models Are Lost Without the Right Context
ProgressCompass supplies missing context so embodied progress reward models estimate long-task progress more accurately.
Progress reward models score how far an embodied task has advanced, but long tasks often cannot be judged from the current frame. ContextProgress-Bench includes 24 manipulation tasks and 120 episodes spanning state recall, sequence tracking, and recurrence disambiguation. Even models that read full history fail without the right context; supplying it cuts error by 77-82% for five models. ProgressCompass, an agentic loop that uses general-purpose vision-language models to supply context, cuts a frozen PRM's error by 63% and raises rank agreement by 76%.
- ContextProgress-Bench has 24 manipulation tasks and 120 episodes.
- Correct context cuts progress error 77-82% for five models.
- ProgressCompass reduces frozen-PRM error 63% and raises rank agreement 76%.
From the original article
1 min read at huggingface.co
Embodied agents now take on ever longer tasks. For long tasks, knowing only whether a task finally succeeds or fails says little; the steps along the way matter. Progress Reward Models (PRMs) score how far a task has come at every step, and serve as dense rewards, verifiers and monitors.
By Jianshu Zhang, Keliang Wu, Chengxuan Qian · huggingface.co
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.