---
library_name: pytorch
language:
  - en
tags:
  - arxiv:2610.02784
  - robotics
  - tactile
  - vision-language-action
  - UniVTAC
---

# SimpleTouch

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

[Project Page](https://simpletouch-robot.github.io/) · [Paper](https://arxiv.org/pdf/2610.02784) · [arXiv](https://arxiv.org/abs/2610.02784) · [Code](https://github.com/ccyydd/SimpleTouch)

SimpleTouch adds a tactile expert to a pretrained π0.5 policy and learns contact-rich manipulation from task demonstrations, without additional tactile policy pretraining. A frozen T3 encoder represents two RGB tactile images, and the action expert reads the tactile expert's layerwise context. Future-touch prediction provides auxiliary supervision during training and is not run at inference.

## Checkpoints

This repository provides task-specific **20k-step inference checkpoints** for six UniVTAC tasks.

| Task | Directory | Actions executed per policy query |
| --- | --- | ---: |
| Lift Bottle | `lift_bottle` | 50 |
| Pull-out Key | `pull_out_key` | 50 |
| Lift Can | `lift_can` | 50 |
| Put Bottle | `put_bottle_in_shelf` | 50 |
| Insert Hole | `insert_hole` | 25 |
| Insert Tube | `insert_tube` | 25 |

These are the default H50/H25 **execution horizons**, not the number of flow-matching inference steps. Temporal ensembling is off. All six checkpoints use the Fast inference path.

Each task directory contains:

```text
<task>/global_step_20000/
├── actor/model_state_dict/full_weights.pt
├── config.yaml
├── norm_stats.json
└── _SUCCESS
```

Use the SimpleTouch loader, which reads the configuration and normalization statistics and validates the inference bundle. Optimizer state and distributed training-resume files are not included.

## Quick Evaluation

Follow the [setup instructions](https://github.com/ccyydd/SimpleTouch/blob/main/README.md) for the policy environment, UniVTAC, T3, and tokenizer first. Run the following commands from the code repository root, with the policy environment active and those paths configured.

**Single GPU:**

```bash
export TASK=insert_hole
GPU_ID=0 bash scripts/simpletouch/evaluate_univtac.sh
```

The launcher automatically downloads and validates the selected checkpoint on first use. Change `TASK` to any directory name in the table above.

**Multiple GPUs:** split the same 100 evaluation episodes across GPUs and merge the results into one directory.

```bash
export TASK=insert_hole
SIMPLETOUCH_EVAL_GPUS=0,1,2,3 bash scripts/simpletouch/evaluate_univtac.sh
```

For a short single-GPU smoke test, prefix the command with `TOTAL_NUM=3`. To use an existing local download, also set `CKPT_PATH=/path/to/insert_hole/global_step_20000`.

Results are saved under `results/UniVTAC/SimpleTac/<task>/<run-id>` in the code repository. Set `SIMPLETOUCH_EVAL_ROOT` to use another output root.

## Dependencies and Scope

- Obtain the [T3-large Mini encoder](https://huggingface.co/datasets/alanz-mit/FoundationTactile/tree/main/models/t3_large) and PaliGemma tokenizer using the instructions in the code repository.
- Install the [UniVTAC](https://github.com/univtac/UniVTAC) simulation environment and assets separately. They are not included in these weight bundles.
- These checkpoints are intended for their corresponding UniVTAC tasks, not direct deployment on physical robots. Third-party models, datasets, and dependencies retain their respective licenses and access requirements.

## Citation

```bibtex
@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},
}
```