Touch meets vision-language-action

SimpleTouch.

Can Vision-Language-Action Models Master
Contact-Rich Manipulation Without
Tactile Policy Pretraining?

A simple extension. A richer sense of contact.
One tactile expert, learned from task demonstrations.

Chen Yang1,2,*Linzhe Shi2,*Changjie Wu2Hang Zhang2Ronghan Chen2Lingjun Zhang1,2Xu Hu2,3Mu Xu2Jiansheng Fan1Chen Wang1,†
1Tsinghua University2Amap, Alibaba Group3Hong Kong Polytechnic University

* Equal contribution · † Corresponding author

A closer look at contactFour real-world tasks
PiPER performing USB insertion with tactile sensors01 / Insert USB
Robot picking up a fragile chip02 / Pick Up Chips
Robot manipulating a Mahjong tile03 / Play Mahjong
Wrist-camera view of the board-wiping task04 / Wipe Board
Explore the film

50Demonstrations per task
10Contact-rich tasks
77.5%UniVTAC average success
71.3%Real-world average success

The central idea

More touch.
Less training overhead.

Strong tactile manipulation does not have to start with another policy-pretraining stage.

SimpleTouch builds on a pretrained π0.5 policy and a pretrained tactile encoder. A tactile expert learns how to use contact information directly from task demonstrations, with future-touch prediction providing additional supervision.

There is no additional tactile policy pretraining and no separate visuotactile alignment stage.

Read the full abstract

The complete abstract is available in the paper.

01 / Method

A small addition.
A richer contact signal.

Keep the pretrained VLA.
Add a tactile expert.
Learn from the same task demonstrations.

01

Preserve the representation

A frozen T3 encoder retains every global and spatial token from current touch.

02

Learn a tactile–action interface

The action expert reads the tactile expert’s layerwise keys and values alongside visual-language context.

03

Learn from future touch

Multi-horizon tactile prediction trains the same expert that supplies current-touch context for action generation.

SimpleTouch architecture. Current touch informs action generation; future tactile latents provide training-only supervision.

Action learning + future-touch prediction

The VLA policy and tactile expert are optimized jointly while T3 stays frozen. Predictive supervision spans multiple future moments, using pooled tactile targets.

One task-training stage

02 / Results

Simple design.
Strong task performance.

Highest success among evaluated methods
across six simulation and four real-world tasks,
with 50 demonstrations per task.

Average success rate (%) ↑

UniVTAC simulation

77.5%

Average task success

+10.8 percentage points over FTP-1

Highest success on every evaluated UniVTAC task, without additional tactile policy pretraining.

Six contact-rich tasks · 50 demonstrations per task · 100 evaluation episodes per task. Avg. w/o Lifts excludes Lift Bottle and Lift Can.

All task-level results
Task success rates (%)

Bold indicates the best result; underlining indicates the second-best result. See the paper for method-specific evaluation protocols.

03 / What makes it work

The details behind the simplicity.

Explore the ablations
Spatial detail

Keep the current
touch intact.

Using all 197 current tactile tokens per sensor achieves 74.3% mean success, compared with 57.5% using only the CLS token.

CLS only57.5%
All tokens74.3%
Temporal supervision

Learn contact
as it changes.

Multi-horizon prediction reaches 74.3%, versus 68.5% for single-horizon prediction at the same per-horizon spatial resolution.

Single-horizon68.5%
Multi-horizon74.3%
Data efficiency

Start with fewer
demonstrations.

On Put Bottle and Insert Hole, SimpleTouch with 20 demonstrations outperforms FTP-π0.5 with 50 demonstrations.

60% fewer

demonstrations at these compared settings

Token and prediction ablations report the mean over Pull-out Key, Put Bottle, Insert Hole, and Insert Tube.

Go deeper

A simpler route to
contact-rich manipulation.

Read the full method, controlled ablations, and evaluation protocols.

THE PAPER SimpleTouch paper overview, including the model and task performanceSimpleTouch

Can Vision-Language-Action Models Master Contact-Rich Manipulation Without Tactile Policy Pretraining?

Citation

@misc{yang2026simpletouchvisionlanguageactionmodelsmaster,
  title={SimpleTouch: Can Vision-Language-Action Models Master Contact-Rich Manipulation Without Tactile Policy Pretraining?},
  author={Chen Yang and Linzhe Shi and Changjie Wu and Hang Zhang and Ronghan Chen and Lingjun Zhang and Xu Hu and Mu Xu and Jiansheng Fan and Chen Wang},
  year={2026},
  eprint={2610.02784},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2610.02784},
}

SimpleTouch architecture

Scroll to inspect the full-resolution figure.