The Ant Lingbo team found through experiments that feeding raw tactile data directly into action models actually lowers robot manipulation success rates. This is because tactile data is sparse and noisy, and data formats vary widely across sensors. To address this, the team has open-sourced LingBot-Touch, a tactile pre-training foundation model that serves as a "translator" between sensors and action models. Through a three-step training process—unifying contact point coordinates, adopting the JEPA approach to predict in feature space, and expanding diverse tactile data—it uniformly encodes signals from different types of tactile sensors into tactile tokens that action models can understand. Empirical tests show that integrating this model substantially improves the manipulation success rates of multiple action models across different dexterous hands, and that it can adapt to sensor hardware not seen during pre-training. The model fills the gap in perception capability after a robot makes contact with an object and represents an important advance in the embodied-native technology route, enabling the reuse of contact experience across different hardware to improve generalization in physical manipulation. The team has already begun related commercialization testing. Released as open source, the model can adapt to a variety of existing action models and sensors, helping to address the industry's challenge of fine manipulation in the physical world.