Event
Tsinghua Team Proposes SimpleTouch Tactile-Enhanced VLA Method
1 reports · 1 sources
A research team at Tsinghua University proposes the SimpleTouch tactile-enhanced VLA method: a tactile expert is added to the pretrained π0.5 model, with single-stage training using only 50 demonstrations per task, requiring no additional tactile policy pretraining or separate alignment stage. The team reports that the method achieves the highest success rate on 6 simulated tasks and 4 real-robot tasks, surpassing FTP-1, which underwent large-scale tactile policy pretraining. In comparisons, all methods used the same 50 task demonstrations, with each simulated task evaluated 100 times. SimpleTouch freezes the pretrained T3 encoder and retains all 197 current tactile tokens per sensor (including the global CLS token and 196 spatial tokens); at deployment, it encodes only the current tactile input and reuses the cache, without running the prediction branch.
Report timeline
- Tsinghua team proposes SimpleTouch: fine manipulation learned from 50 demonstrations without tactile policy pretraining
A Tsinghua University research team proposes SimpleTouch, which adds a tactile expert to the pretrained VLA π0.5 and uses single-stage training with only 50 demonstrations per task, without additional tactile policy pretraining or a separate alignment stage.
Xbotics具身智能实验室 · 微信公众号 ·