Paper review · November 11, 2026
Learning tactile representations across different tactile sensors for general-purpose dexterous manipulation
Learning Heterogeneous Tactile Representation for General Dexterous Manipulation
The authors built LingBot-Touch, a tactile foundation model that combines data from piezoresistive, magnetic and visuotactile sensors into a single representation, and used it to improve the dexterous manipulation performance of several policy models.
Problem
Commercial tactile sensors each have clear strengths and weaknesses and different physical properties, so tactile understanding learned with one sensor is hard to transfer to another. This makes it difficult to use heterogeneous tactile data collected from multiple sensors together.
Method
The authors developed LingBot-Touch, a tactile foundation model that unifies piezoresistive, magnetic and visuotactile sensors in a single representation space. Its output tactile tokens are plug-and-play, so they can be attached directly to Diffusion Policy or to VLA models such as π0.5 and LingBot-VLA. The model is also designed to extend easily to new sensor types.
Results
- The authors report that, in simulation and in real-robot experiments on several hardware platforms, this tactile representation substantially improved performance on complex manipulation tasks.
- They report that adding the tactile tokens improved manipulation performance across different policy models, from Diffusion Policy to π0.5 and LingBot-VLA.
- They state that computational efficiency also stays high (the abstract gives no specific figures).
Limitations
This analysis is based on the abstract. The abstract gives no specific figures, baselines or list of evaluation tasks, so the size of the improvement cannot be checked. Only three sensor types are supported, so whether the approach extends to other sensors needs further validation.
Why it matters
It suggests that tactile learning can be scaled up by pooling data from multiple sensors instead of building a separate model for each sensor. It could be a practical way to add touch to existing VLA models.
Key terms
- Visuotactile sensor
- A sensor that obtains tactile information by using a camera to capture how a contact surface deforms when pressed. It has high resolution, but its data format differs greatly from that of other tactile sensors.
- VLA (Vision-Language-Action) model
- A model that takes images and language instructions as input and directly outputs robot actions. π0.5 is a well-known example.
- Diffusion Policy
- A policy learning method that uses a diffusion model to generate sequences of robot actions.
Original
Authors: Bing Han, Yichuan Peng, Haoyu Dong, Yuanmeng Hu, Yongyi Jia, Fangjing Wang, Kecheng Zheng, Yiming Zhang, et al. · Posted 2026-11-09 · License CC BY 4.0 · Based on the abstract
This review is an AI-written summary of the paper. Please treat the original paper as the reference.