Report: Robbyant Open-Sources Tactile Foundation Model LingBot-Touch

roboticsandautomationnews.com reported on Nov. 15, 2026, that Robbyant, a unit of Ant Group, had open-sourced its tactile foundation model, LingBot-Touch. According to the article, the model was pretrained on about 5,000 hours of real-world tactile data and supports three types of sensing technology: magnetic, vision-based and piezoresistive. shattered.io reported that the model was released on Nov. 10, 2026.

Robbyant reportedly released LingBot-Touch as a model that maps signals from different tactile sensors to a common representation, allowing multiple dexterous hands and action models to share tactile understanding. The design addresses robotics development involving sensors that produce different types of signals, but the reported information does not establish how well the model transfers across different hardware or performs in real-world tasks.

The notable aspect of the announcement is the combination of about 5,000 hours of real-world data and support for three types of sensor technology. Tactile sensors are reported to produce different raw signals depending on their technology, and LingBot-Touch’s common-representation design appears intended to shift from building separate tactile-understanding systems for each sensor to sharing a representation. However, the reported information alone does not show how much sensor variation or differences in contact conditions the common representation can accommodate. We covered the model in “Robbyant’s LingBot-Touch Tactile Foundation Model Introduced by humanoid.guide”. The main source for this report also covers LingBot-Touch’s release, so it should be viewed as an effort to share tactile representations, not as an announcement of a separate model. We covered Robbyant’s embodied AI challenge in “Robbyant Embodied AI Challenge Draws More Than 1,800 Teams From 13 Countries”. The challenge is based on LingBot-VLA 2.0, but the information provided does not establish a technical connection to LingBot-Touch or plans for participants to use the model. The business focus is on treating touch not as something confined to a single hand or action model, but as a representation that can be reused across multiple devices and models. If successful, this could reduce dependence on specific sensors when translating tactile input into fine manipulation. However, reports do not cite results demonstrating real-world task performance or transfer to other hardware. Open-source release points toward broader availability, but ease of adoption cannot be assessed without details about the license, supported hardware and scope of the released implementation. The next questions are what sensors, objects, contact conditions and tasks are represented in the approximately 5,000 hours of data, and how well the common representation can be reused with different hands and action models. The main source cites examples such as turning pages and opening bottle caps, but gives no figures for success rates or safety in physical deployments. The generality of a tactile model depends not only on the size of its training dataset, but also on whether its performance can be reproduced with unfamiliar sensors and manipulation conditions.