TOPReward uses VLM token probabilities for robot learning rewards

Researchers have developed TOPReward, a novel method for generating dense, instruction-conditioned feedback for robotic learning without requiring manual annotations or task-specific reward models. This approach leverages the internal token probabilities of pretrained Video-Language Models (VLMs) to measure task progress, effectively converting latent understanding into a usable reward signal. TOPReward has demonstrated strong performance on real-world manipulation benchmarks and Open X-Embodiment datasets, outperforming other training-free VLM reward methods and showing competitiveness with trained reward models. AI

IMPACT Enables more efficient and scalable robotic learning by providing dense, automated reward signals.

RANK_REASON Academic paper detailing a new method for robotic learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• DagsHub
• Gotit.pub
• Hugging Face
• ManiRewardBench
• ScienceCast
• Shirui Chen
• TOPReward
• Video-Language Models

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →