Robbyant releases LingBot-Touch to help robot hands share tactile understanding
Robbyant, Ant Group’s embodied-AI company, announced LingBot-Touch on November 10, offering a tactile perception layer intended to work across different robot hands and control models. The release targets an integration problem: touch sensors based on different physical principles produce incompatible signals, making experience difficult to reuse. Company announcement
The model converts magnetic, piezoresistive and vision-based tactile measurements into a common representation. Its research paper, submitted November 9 at 18:12 UTC, describes integration with diffusion policies and vision-language-action models rather than replacing those controllers. Technical preprint
Robbyant says the pretraining corpus contains 4,983 hours of tactile recordings. Its project page describes tests on 26 real-world tasks using Sharpa Wave and XHand hardware, with 20 trials per task. These are developer-reported experiments, not independent demonstrations of reliability across arbitrary robots. Project and evaluation details
The company reports average task-success gains of roughly 14 to 26 percentage points across four action-model families relative to runs without tactile input. That comparison matters: improvements over no touch should not be read as the same gain over competing tactile encoders. Company results
The public GitHub repository contains an inference implementation, a frozen encoder interface, sample data and links to Tiny and Small checkpoints. It labels the release Apache-2.0, while noting that model files are hosted separately and require an account with repository permission. Unrestricted weight access and checkpoint-specific terms have not been independently verified here. Repository, code license
For robotics developers, the immediate opportunity is a reusable touch-feature extractor. It is not a complete autonomous manipulation system: sensor geometry, signal normalization and downstream policy integration still matter. No model execution or independent replication was performed for this report.
Sources and citations
Revision history
Published .