Pretrained vision-language-action policies carry strong language-conditioned manipulation priors, but they remain largely vision-driven. In contact-rich states, visual observations can be ambiguous: occlusion, partial observation, and depth error make it hard to know whether the robot has actually grasped, inserted, or seated an object.
LIFT (Late Reactive Injection of Force for VLA Post-Training) adds force after pretraining instead of rebuilding the foundation model. It encodes recent 6D end-effector force as causal force memory, injects it through zero-initialized cross attention, and trains a reactive action expert beside the original VLA action stream.
Across towel folding, book insertion, and Hanoi ring placement, LIFT learns faster and reaches stronger performance than vision-only post-training while preserving the pretrained VLA prior.
LIFT starts from a pretrained vision-only VLA and grafts a reactive action expert beside the original action expert. Instead of committing to a full open-loop action chunk, the reactive stream decodes actions causally within the chunk and receives recent 6D end-effector force as latency-aligned force memory.
At runtime, the slow vision-language prefix is computed once and cached. Each within-chunk refresh only re-encodes the newest force window and reruns the lightweight reactive branch, allowing the policy to respond to contact changes without a full vision-language forward pass.
Adding a force pathway should not erase the pretrained VLA prior. LIFT copies the original action-expert weights into the reactive expert and uses a shifted causal attention pattern so each reactive token receives an equivalent context to the original fully attentive action token at initialization.
The force-injected cross attention is added as a residual update with a zero-initialized output projection. Before post-training, the force residual is exactly zero, so the augmented model starts from the same action output as the base VLA.
LIFT trains one model on both abundant vision-only task-alignment data and scarce force-enabled online corrections. Vision-only batches use zero force placeholders and mask the encoded force memory, while online correction batches keep measured force active.
The post-training loop mixes offline visual data with online DAgger corrections. This lets the policy preserve broad task knowledge while learning force responses from states visited by the current policy.
We evaluate LIFT on three real-robot contact-rich tasks: towel folding, book insertion, and Hanoi ring placement. The comparison includes LIFT, the full reactive force-injection system; LIFT w/o Reactive Force Injection, which uses force without the reactive force-memory pathway; LIFT w/o Online DAgger, which removes repeated on-policy corrective updates; and π0.5 w/ Online DAgger, a vision-only online post-training baseline.
LIFT
LIFT
π0.5 w/ Online DAgger
π0.5 w/ Online DAgger
LIFT
LIFT
π0.5 w/ Online DAgger
π0.5 w/ Online DAgger
LIFT w/o Reactive Force Injection
LIFT w/o Reactive Force Injection
LIFT
LIFT
π0.5 w/ Online DAgger
π0.5 w/ Online DAgger
LIFT w/o Reactive Force Injection
LIFT w/o Reactive Force Injection
This question compares LIFT and LIFT w/o Reactive Force Injection against π0.5 w/ Online DAgger. The result shows that force-aware post-training learns faster and reaches stronger task performance than vision-only online post-training.
This question compares LIFT with LIFT w/o Reactive Force Injection. The result shows that recent force memory and within-chunk reactive updates are important for contact phases that cannot be resolved from a single force frame.
This question compares LIFT with LIFT w/o Online DAgger. The result shows that offline force data alone misses many contact states encountered by the learner, while online DAgger collects corrections on the policy's own failure states.
This question evaluates the final LIFT checkpoint under object, tablecloth, and lighting shifts. The result shows that LIFT keeps the visual generalization inherited from the pretrained VLA while adding force-aware contact adaptation.
@article{wang2026lift,
title = {LIFT: Never Too Late for Force -- Accelerating VLA Post-Training with Reactive Force Injection},
author = {Wang, Yi and Chen, Wendi and Wen, Zimo and Xue, Han and Li, Xueqi and Yu, Wenye and Chen, Zhijie and Yang, Hao and Lv, Jun and Wen, Chuan and Lu, Cewu},
journal = {arXiv preprint arXiv:2607.14236},
year = {2026}
}