Venue: ICML 2026 (Poster) Category: Tactile Affiliations: Authors: Yan Zhang, Zheng Wang, Pengpeng Zeng, Xing Xu, Jingkuan Song, Heng Tao Shen

Visuo-tactile sensors are now widely used in robotic manipulation, but the "inherent heterogeneity in sensor designs hinders the learning of unified tactile representations in cross-sensor scenarios." Existing methods that rely on reconstruction or task-specific supervision "often fail to capture the common information between different tactile sensors, particularly in the presence of substantial sensor variations, resulting in limited generalization to unseen sensors."

The authors propose Cross-Tactile Sensor Representation Learning (CTSRL), a unified framework for sensor-agnostic tactile representation learning. Its key elements:

• Cross-Sensor Modulator (CSM) to "eliminate sensor-specific biases," so downstream representations are not entangled with the idiosyncrasies of any one sensor design.
• Two-stage learning paradigm:
  1. Synthetic stage — "leveraging aligned synthetic data for cross-sensor self-supervised learning to extract shared latent representations across sensor domains."
  2. Real stage — "integrating real-world multimodal tactile data to bridge the sim-to-real semantic gap through cross-modal alignment, thereby enriching representations with fine-grained semantic attributes."

flowchart TD
    S1[Aligned synthetic tactile data] --> CSM[Cross-Sensor Modulator<br/>remove sensor-specific bias]
    CSM --> SSL[Stage 1: cross-sensor<br/>self-supervised learning]
    SSL --> SHARED[Shared latent representation]
    S2[Real-world multimodal tactile data] --> CMA[Stage 2: cross-modal alignment<br/>bridge sim-to-real gap]
    SHARED --> CMA
    CMA --> REP[Sensor-agnostic representation<br/>fine-grained semantics]
    REP --> GEN[Generalizes to unseen sensors]

The available abstract reports no specific numerical results. It states that "our method demonstrates strong multi-sensor generalization, significantly improving sensor-agnostic representation learning" relative to prior reconstruction- or task-supervision-based approaches.

The tactile-sensing field is fragmented across many incompatible sensor designs (GelSight, DIGIT, and others), and models rarely transfer between them. CTSRL's two-stage synthetic-then-real recipe plus an explicit sensor-bias modulator is a step toward reusable, sensor-agnostic tactile foundations — a prerequisite for tactile data and policies to be shared across the diverse hardware in the 2026 manipulation landscape.

• ICML 2026: https://icml.cc/virtual/2026/poster/66793

← Back to ICML-2026