LIFT: Teaching VLA Models Force Perception via Post-Training Without Force Pretraining

Shanghai Jiao Tong University researchers (Lu Cewu and Wen Chuan teams) with Qongche propose LIFT (Late Reactive Injection of Force for VLA Post-Training), a method that teaches Vision-Language-Action (VLA) models force perception solely through post-training, without any force-labeled pretraining data.

Problem: Force Data Scarcity and Modality Integration

Force sensing is critical for contact-rich manipulation (e.g., book insertion, ring placement) where visual cues barely change during contact. However, force data collection is expensive and tightly bound to specific hardware, making large-scale pretraining impractical. Adding a new modality also risks catastrophic forgetting of pretrained spatial reasoning and generalization.

LIFT Approach

Reactive Force Injection for High-Frequency Control

LIFT duplicates the pretrained action expert into a "reactive action expert" that attends to recent force history via causal cross-attention, adjusting actions at 10 Hz while vision-language context is precomputed and cached (up from 1 Hz).

Preserving Pretrained Knowledge via Weight Copying and Shifted Causal Attention

At initialization, the reactive expert's weights are copied from the base expert. A shifted causal attention mask ensures the reactive tokens receive identical context as base tokens, aligning outputs numerically and preserving priors.

Two-Stage Training with Mixed Visual and Force Data

Stage 1: Collect force-free visual data using RoboPocket to learn task basics. Stage 2: Deploy on robot, gather 20–30 force-labeled corrections via human intervention. During stage 2, visual-only samples have their force gradients masked to avoid overfitting.

Experimental Results

Evaluated on three tasks:

Towel folding: 73.3 → 84.2 success rate

Book insertion: 36.7 → 58.3

Hanoi ring placement: 26.7 → 56.7

Only 20–30 online force episodes (2–3 hours) were needed. Force improved even non-contact-heavy tasks (towel folding) and accelerated learning.

Four Key Conclusions

Post-training alone suffices for force learning – no force-pretraining required.

Reactive injection is critical – enables within-chunk corrections; especially vital for dense-contact tasks.

Online data collection matches policy distribution – mitigates force distribution shift caused by minor position changes.

Pretrained knowledge is retained – generalization to cloth, lighting, and object changes remains intact after post-training.

Code and details: https://arxiv.org/abs/2607.14236, https://lift-policy.github.io/, https://github.com/y-wng/lift.