Ant lion bot releases LingBot-Touch, open-sourcing tactile foundation model to enhance robot dexterity

Ant Lion Bot has publicly released LingBot-Touch, a tactile foundation model designed to equip robotic systems with crucial haptic perception capabilities. Pre-trained on nearly 5,000 hours of real tactile data, LingBot-Touch integrates three major sensing principles—magnetic, visual-tactile, and piezoresistive—to translate diverse signals into a unified tactile representation, enabling cross-sensor and cross-model reuse of accumulated 'feelings'. Testing across 26 real-world scenarios using two dexterous hands, Sharpa Wave and XHand, demonstrated significant performance gains. For instance, in a bottle cap twisting task using the T-Rex controller on the Sharpa Wave, success rates jumped from 56.25% without tactile input to 90% with the LingBot-Touch representation. The model improved the average task success rate across four action models (LingBot-VLA, $\pi$0.5, Diffusion Policy, and T-Rex) by 14 to 26 percentage points. This architecture allows for data from human-worn gloves and remote dexterous hands to be used for training, while also supporting future sensor integration. The company provided two versions, with the lightweight LingBot-Touch-Tiny (around 5.3 million parameters) used in the tests, adding only 0.13% to 0.16% to the action model parameters. Ant Lion Bot asserts that this release underscores a commitment to the embodied AI paradigm, recognizing that high-quality sensory input is vital for physical interaction.