Ant Lingbo reportedly open-sources tactile model “LingBot-Touch”

zhidx.com reported on November 10, 2026, that Ant Lingbo had open-sourced its tactile model “LingBot-Touch.”

zhidx.com reported that the model was open-sourced. If the release’s specifications and terms of use can be confirmed, outside developers may be able to participate in its evaluation. Reproducible evaluation conditions and results from testing on different robots will be important to determining its practical significance.

In discussions of AI for robots, a key issue is not only model performance but also how training data is collected and handled, and whether it can be made available for validation on other systems. This report concerns the release of a tactile model, but this brief does not repeat its specific specifications. Open-sourcing may make it possible for others to validate the research, but the release will become a basis for comparison among developers only if its terms of use, implementation and evaluation procedures are also shared. In a related article, NVIDIA announces GR00T 2 and Isaac Lab 3.0 at CoRL, we reported that GR00T 2 was trained on data that included 100,000 hours of first-person-view footage, and that simulation training is expanding from walking to precision assembly tasks. This reflects a move to connect model training with simulation. By contrast, Maxinsights raises $30 million to expand data infrastructure for Physical AI described the collection of more than 2 million hours of real-world activity footage, with more than 1.5 million hours made available for model use, and introduced “MaxGlove,” a device for capturing tactile information. The report can be viewed in the context of the chain linking data collection and preparation, training, and validation in simulation or on physical robots: where the release of a model that handles tactile information fits into that chain is a key question. However, data volume or whether a model is open source alone does not establish transferability between robots or effectiveness in real-world tasks. Points to watch include whether third parties can reproduce results under the same conditions, whether the model can be evaluated with different sensors and robots, and whether its inputs, outputs and terms of use are clear. If this information is provided, it will be easier to distinguish learning tied to a specific robot from learning assets that can be shared across multiple platforms. This article does not cover the report’s specific specifications or testing, so readers should consult the original source for details.