New CLIFT method enables closed-loop fine-tuning for Gemini Robotics On-Device

Researchers have developed CLIFT, a novel method for fine-tuning robot foundation models, specifically applied to Gemini Robotics On-Device (GROD). This technique allows for closed-loop policy improvement using only API-compatible supervised data, bypassing the need for direct access to model weights or gradients. Through iterative fine-tuning with reward feedback, CLIFT significantly enhances GROD's performance on complex humanoid manipulation tasks, achieving near-perfect success rates. AI

IMPACT Enables more effective adaptation of closed-weight robot foundation models for specialized tasks without direct weight access.

RANK_REASON Academic paper detailing a new method for robot fine-tuning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Gemini Robotics On-Device
• GROD

• model release
• product
• other

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