LIFT: 20-30 Force Traces Teach Pretrained VLA Models Reactive Force Control

机器之心 · wechat · 2026-09-27

A Shanghai Jiao Tong University team (Lu Cewu, Wen Chuan) with QunChu proposed LIFT (Late Reactive Injection of Force for VLA Post-Training), accepted at CoRL 2026 with high scores and open-sourced.

LIFT teaches pretrained VLA models force perception purely via online post-training—no force-data pretraining needed. On towel folding, book insertion, and Hananoi tasks it significantly outperforms vision-only post-training (e.g., 26.7→56.7) with only 20-30 online force traces. It preserves pretrained knowledge via action-expert weight copying and shifted causal attention masks, raises control-loop frequency from 1Hz to 10Hz, and mixes in force-free data to avoid overfitting. The authors argue the paradigm applies to any MoT-based pretrained model.

Original post →

More from Embodied

Embodied channel →