CLIFT pushes Gemini Robotics On-Device to near-perfect humanoid task success
A no-weight-access method hits near-perfect humanoid manipulation in just two cycles
Robotics foundation models are powerful but often closed-source, and managed supervised fine-tuning (SFT) APIs—where users submit data and get a tuned policy—restrict improvement to pure imitation. For agile, contact-rich humanoid manipulation, the gap between imitation outputs and real-world deployment is large. In a new arXiv paper, UC Berkeley researchers present CLIFT (Closed-Loop Iterative Fine-Tuning), a technique that works within the managed-API regime to close this gap using deployment-time reward feedback.
Using Google's Gemini Robotics On-Device (GROD), the team ran direct SFT through the API and found it substantially outperformed a leading open-weight vision-language-action model on the same demonstrations, but still fell short of deployment-level mastery. CLIFT then transforms reward signals from real robot rollouts into API-compatible supervised data, enabling iterative closed-loop improvement without ever accessing weights, gradients, likelihoods, or losses. After just two flywheel cycles, GROD achieved near-perfect success on agile, contact-rich tasks—demonstrating that closed-loop policy improvement is possible even within restrictive managed-API training regimes.
- CLIFT achieves near-perfect success on humanoid tasks after 2 flywheel cycles using Google's GROD model
- Direct SFT via managed API outperforms leading open-weight VLAs but fails on agile manip tasks
- Method converts deployment-time reward feedback into supervised data without accessing weights, gradients, or losses
Why It Matters
Enables closed-loop robot learning on proprietary models, unlocking humanoid adaptation without weight access.