Model family

LingBot-Touch

LingBot-Touch is Robbyant’s JEPA-based tactile representation model. A shared Transformer unifies magnetic, piezoresistive and decoded visuotactile signals into addressed latent tokens for downstream manipulation policies. Inference code and two encoder checkpoints are available.

Overview

Model type
Component
Architecture
Transformer
Listed versions
LingBot-Touch 1.0
Family introduction

Model history

November 2026

LingBot-Touch introduces a shared encoder for heterogeneous tactile sensing

New modelVendor reported

The November9 v1 preprint introduces LingBot-Touch: modality-specific tokenizers and one JEPA-trained Haptic Transformer produce reusable tactile tokens across piezoresistive,magnetic and decoded visuotactile inputs. Frozen encoders are studied with downstream policies on real hands and in simulation. The model represents touch; it does not itself output robot actions. The paper announces planned Tiny/Small inference resources,so its submission date is not treated as an independently verified first-weight availability date.

Previously reported

Tactile recordings differ in sensing physics, channels and taxel layouts; the report identifies sensor-specific representations and adaptation requirements as obstacles to sharing experience across robot hands. This is the paper’s technical baseline, not a claim that no prior cross-sensor model existed.

This announcement

The November9 v1 preprint introduces LingBot-Touch: modality-specific tokenizers and one JEPA-trained Haptic Transformer produce reusable tactile tokens across piezoresistive,magnetic and decoded visuotactile inputs. Frozen encoders are studied with downstream policies on real hands and in simulation. The model represents touch; it does not itself output robot actions. The paper announces planned Tiny/Small inference resources,so its submission date is not treated as an independently verified first-weight availability date.

Featured because this named tactile foundation-model introduction materially expands the assessed ability to share contact representations across sensor types and robot hands. The significance is the new model and its robot-learning role,not the admission score,a minor paper revision or an asserted universal performance breakthrough. Evaluations remain developer-reported; current downloadable weights are verified separately.

View all important updates →

Papers & reports (1)

Learning Heterogeneous Tactile Representation for General Dexterous Manipulation

Paper available

arXiv: 2611.12806DOI: 10.48550/arxiv.2611.12806
Authors & affiliations (19)

Author order: Original order

  1. Bing Han

    Not provided

  2. Yichuan Peng

    Not provided

  3. Haoyu Dong

    Not provided

  4. Yuanmeng Hu

    Not provided

  5. Yongyi Jia

    Not provided

  6. Fangjing Wang

    Not provided

  7. Kecheng Zheng

    Not provided

  8. Yiming Zhang

    Not provided

  9. Yanhong Zeng

    Not provided

First published
2026-11-09
Latest paper revision
v1 · 2026-11-09
Venue
Not provided

Covered model versions

  • lingbot-touch · LingBot-Touch 1.0

    LingBot-Touch1.0 tactile representation encoders; Tiny and Small, frozen-encoder probes and policy-conditioned robot/simulation studies. Paper result revisions are not equated with current native bundle SHAs.

Paper ↗PDF ↗

Official resources

Submit a source or correction