CLIFT Enables Closed-Loop Fine-Tuning for Closed-Weight Robotics Models
August 3, 2026
CLIFT provides a method to turn closed-weight robot foundation models, like Gemini Robotics, into humanoid specialists using non-invasive, iterative fine-tuning. It bypasses the limitations of supervised fine-tuning APIs that restrict users to pure imitation, allowing for closed-loop policy improvement in contact-rich manipulation tasks.
HOW THIS AFFECTS YOU
●
builderYou can now improve proprietary robot policies using closed-loop methods instead of being limited to imitation learning.
●
researcherThis provides a workaround for training reinforcement learning policies on closed-weight models.