LingBot-Touch: Unified AI Model for Heterogeneous Tactile Sensors

Researchers have unveiled LingBot-Touch, a groundbreaking tactile foundation model designed to unify and extend tactile representation across heterogeneous sensors. Published on arXiv on November 9, 2026, and featured on HuggingFace Daily Papers the following day, the model addresses key challenges in robotic manipulation by bridging sensor-specific physics and enabling seamless integration across diverse tactile systems.

Unified Representation Across Tactile Sensors

LingBot-Touch employs a JEPA-based architecture built around a shared Transformer and latent representation prediction from spatiotemporal tactile contexts. This design allows the model to process data from piezoresistive, magnetic, and visuotactile sensors within a unified latent space. The resulting tactile tokens are plug-and-play, enabling integration into policy models such as diffusion policy and vision-language-action (VLA) frameworks, including π_0.5 and LingBot-VLA.

Key Properties and Performance

The research highlights four critical properties of LingBot-Touch:

• Unified: Heterogeneous tactile data mutually enhance representation learning.
• Transferable: The model transfers seamlessly to new sensors and robotic platforms without additional training.
• Efficient: High-frequency downstream inference is supported even with a lightweight Vision Transformer (ViT) backbone.
• Scalable: Performance improves as the volume of tactile training data increases.

Extensive experiments, including simulations and real-robot deployments, demonstrate that LingBot-Touch enhances dexterous manipulation tasks while maintaining computational efficiency. The model's scalability and adaptability make it a promising tool for advancing tactile learning and robotic hardware design.

Open-Source Availability

The LingBot-Touch implementation is open-source and available on GitHub under the Apache License 2.0. The repository has garnered 103 stars and supports high-frequency inference and cross-task transfer, making it accessible to researchers and developers aiming to leverage tactile data in robotics.

This development underscores the potential of integrating cross-sensor tactile data to scale embodied intelligence, offering valuable insights for future advancements in robotics and sensor technology.