LIFT gives robots reactive force feedback for faster, better contact-rich manipulation

Pretrained vision-language-action (VLA) policies excel at language-conditioned manipulation but are fundamentally vision-driven, struggling once contact occurs—occlusions, ambiguous depth, and small force errors derail execution. To solve this, researchers from Shanghai Jiao Tong University and collaborators propose LIFT (Late Reactive Injection of Force for VLA Post-Training). LIFT grafts a reactive action expert beside the original one, initializing it from pretrained action weights. It injects recent 6D end-effector force through a causal force memory and zero-initialized cross-attention, allowing actions to be refreshed mid-execution. To handle the policy-dependent distribution shift of force feedback, LIFT couples reactive injection with an online DAgger loop that trains on a mixture of offline task-alignment data and human-corrected online rollouts.

Across three challenging tasks—towel folding, book insertion, and Hanoi ring placement—LIFT achieves faster learning and higher success rates than vision-only post-training baselines. Ablation studies confirm that both the reactive force memory and the online corrective data are critical for robust performance in contact-rich scenarios. The method preserves the VLA policy's general manipulation knowledge while adding essential contact reactivity. The team will release code and data, offering a practical path to make pretrained VLA policies truly capable of handling the physical world where touch matters as much as vision.