Pretrained vision-language-action (VLA) policies provide strong language-conditioned manipulation knowledge but struggle in contact states where occlusion, depth ambiguity, or small force errors occur.
We introduce LIFT (Late Reactive Injection of Force for VLA Post-Training), a force-aware post-training framework that enhances a pretrained VLA policy with contact reactivity while preserving its general manipulation knowledge.
LIFT grafts a reactive action expert beside the original action expert, initializing it from pretrained action weights, and injects recent 6D end-effector force through causal force memory and zero-initialized cross attention, allowing actions to be refreshed during execution.
To tackle the policy-dependent distribution shift of contact feedback, LIFT couples reactive force injection with an online DAgger loop that trains on a mixture of offline task-alignment data and human-corrected online rollouts.
In tasks like towel folding, book insertion, and Hanoi ring placement, LIFT learns faster and achieves higher performance than vision-only post-training. Ablation studies indicate that both reactive force memory and online corrective data are crucial for robust contact-rich manipulation. Our code and data will be publicly available.
Blogger's Review: The LIFT framework significantly enhances the performance of VLA policies in complex contact tasks by incorporating force feedback. This innovative post-training approach showcases the effective combination of offline and online learning, expanding the potential applications of machine learning in robotic manipulation. This method is worth exploring for researchers and developers alike.