By RXA AI News Desk
November 11, 2026
Robbyant, the advanced robotics and physical intelligence division of Ant Group, has officially open-sourced LingBot-Touch, the world's first generalized touch foundation model built exclusively for humanoid manipulation. The model arrives with over 5,000 hours of high-fidelity tactile sensor data gathered across a vast spectrum of physical handling environments. By employing a Joint Embedding Predictive Architecture (JEPA), LingBot-Touch transforms raw hardware signals into universal tactile tokens, allowing robotic manipulators to instantly recognize texture, slippage, and weight boundaries.
What makes LingBot-Touch a milestone for developers is its hardware-agnostic encoder framework. Instead of forcing teams to rewrite or retrain sensory layers for specific proprietary robotic arms, the encoder seamlessly maps touch tokens to pre-existing visual and motor policies. Roboticists can clone and deploy the repository directly via GitHub to run entirely on local hardware, bypassing the latency boundaries of cloud-hosted robotics platforms.