Berkeley paper turns Gemini Robotics On-Device into a humanoid specialist via CLIFT

berkeley_ai · x · 2026-08-04

Berkeley researchers introduce CLIFT (Closed-Loop Iterative Fine-Tuning), a framework that turns Gemini Robotics On-Device into a humanoid task specialist using only a managed SFT API.
The core challenge is that the proprietary model exposes no weights, gradients, losses, or action log-probabilities, so standard RL methods like PPO cannot be applied directly.
CLIFT encodes the reinforcement signal into supervised training data, letting repeated closed-loop improvement happen through the SFT interface without changing the underlying model or training procedure.
The paper focuses on the deployment gap for closed-weight robot foundation models: they are strong generalists, but still fail on agile, contact-rich humanoid tasks in real rollouts.

More from Embodied