Robots gain shared sense of touch, improved dexterity with China’s 5,000-hour dataset

Robbyant, an embodied AI company under Ant Group, has open-sourced LingBot-Touch, a tactile foundation model designed to give robots advanced touch perception. Pre-trained on nearly 5,000 hours of real-world tactile data, the model supports three sensing modalities: magnetic tactile, visual tactile, and piezoresistive sensing. It converts signals from these different sensors into a unified tactile representation, allowing different robotic hands and action models to share and reuse touch-related information. The company says this capability could improve fine manipulation tasks, including turning pages, unscrewing bottle caps, and identifying Mahjong tiles. Unified robotic touch LingBot-Touch is an open-source tactile foundation model designed to help robots interpret physical contact across different types of sensors. The model addresses a major challenge in robotic manipulation: enabling robots to retain and reuse touch perception capabilities when switching between sensors, dexterous hands, and control models. Commercial tactile sensors currently use different technologies, including magnetic tactile, visual tactile, and piezoresistive sensing. Each approach has distinct advantages, limitations, costs, and application scenarios, making it unlikely that a single sensing technology will dominate the market. However, these different sensing principles generate incompatible raw signals, even when detecting the same physical contact. Models trained to interpret data from one sensor type may struggle to transfer that knowledge to another, requiring developers to collect new datasets, retrain models, and adapt their systems whenever the hardware changes. LingBot-Touch attempts to overcome this limitation by translating heterogeneous tactile signals into a shared representation. Its architecture uses separate encoding pathways to process data from different sensing technologies before mapping contact points from different sensor layouts into a common fingertip coordinate space. A shared model then learns to interpret contact locations, force distributions, and how these properties change over time. The resulting unified tactile tokens allow different robotic hands and action models to access and reuse the same learned touch information. The architecture also allows data collected from humans wearing tactile gloves and from teleoperated robotic hands to be combined during training. This could help developers transfer tactile knowledge across hardware platforms without having to build separate perception systems for each sensor configuration. The company says the architecture can also accommodate new tactile sensing technologies as they emerge. Master tactile skills Robbyant evaluated the model through 26 real-world manipulation tests involving two dexterous robotic hands, Sharpa Wave and XHand. Four action models—LingBot-VLA, π0.5, Diffusion Policy, and TRex—were tested with LingBot-Touch, achieving average task success rate improvements of approximately 14 to 26 percentage points compared with operating without tactile input. One test involved using the T-Rex action model to control Sharpa Wave while unscrewing a bottle cap. The robot achieved a success rate of 56.25 percent without tactile input, increasing to 62.5 percent when raw sensor signals were fed directly into the action model. When LingBot-Touch processed the tactile information before passing it to the policy, the success rate reached 90 percent. The team also tested whether the model could transfer its capabilities to different hardware. XHand uses magnetic tactile sensors that were not included in LingBot-Touch’s pre-training data. Instead, the model learned magnetic tactile information from human-collected glove data. Robbyant kept the foundation model’s parameters unchanged and trained only the action model using XHand demonstration data to evaluate cross-hardware reuse. The company offers two model sizes to balance perception capabilities and computing requirements. All 26 real-world tests used LingBot-Touch-Tiny, which contains approximately 5.3 million parameters. When integrated with LingBot-VLA, π0.5, and T-Rex, it increases their total model size by only around 0.13 to 0.16 percent. Robbyant says the larger model also demonstrates potential for further improvements as its pre-training dataset approaches 5,000 hours. The technical report and source code are available through Hugging Face and GitHub. Get the latest in engineering, tech, space & science - delivered daily to your inbox.Jijo is an automotive and business journalist based in India. Armed with a BA in History (Honors) from St. Stephen's College, Delhi University, and a PG diploma in Journalism from the Indian Institute of Mass Communication, Delhi, he has worked for news agencies, national newspapers, and automotive magazines. In his spare time, he likes to go off-roading, engage in political discourse, travel, and teach languages.