New PRIMO R1 framework turns AI into active critics for robotic manipulation

Researchers have developed PRIMO R1, a 7B framework that enhances robotic manipulation by transforming video MLLMs from passive observers into active critics. This system uses reinforcement learning to encourage explicit Chain-of-Thought generation for progress estimation, anchored by initial and current state images. Experiments show PRIMO R1 achieves state-of-the-art performance, reducing mean absolute error by 50% compared to specialized reasoning baselines and outperforming larger general MLLMs. It also demonstrates strong zero-shot generalization on failure detection tasks, surpassing models like OpenAI o1 on the RoboFail benchmark. AI

IMPACT Enhances robotic manipulation capabilities by enabling AI to actively assess task progress and detect failures.

RANK_REASON Academic paper detailing a new AI framework and benchmark. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• OpenAI
• Openai O1 System Card
• PRIMO Dataset
• PRIMO R1
• reinforcement learning
• RoboFail
• supervised fine-tuning
• Yibin Liu

• paper
• model release
• product

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