The XHand1 robot hand’s average contact-center error dropped to 0.16 mm when utilizing Robbyant’s LingBot-Touch model, compared to 5.89 mm with a device-specific adapter. This result demonstrates that the open-source tactile AI can generalize its sensing capabilities to robot hardware not encountered during its training. In real-world tests across six tasks, the robot’s success rate increased to 75.04% with the model’s features, up from 57.21% using raw tactile input or no touch data.
The foundation model uses three 5.46M parameter encoders based on a Joint-Embedding Predictive Architecture and was pretrained on nearly 5,000 hours of real data from magnetic, piezoresistive, and visuotactile sensors. The research team from Ant Group reports that this latent prediction approach handles the sparse and noisy nature of tactile data better than masked autoencoders. In simulated plate spinning trials, the model raised success rates from 0.5% to 97.4% and is released under an Apache 2.0 license.
