LingBot-Touch Brings JEPA To Robotics

The compelling idea is treating touch as a transferable representation-learning problem rather than a hardware-specific signal-processing task.

• –Nearly 5,000 hours of heterogeneous tactile data feed one shared latent space
• –Canonical geometry and modality-specific tokenizers let different sensors communicate through one Transformer
• –Latent prediction avoids wasting capacity reconstructing sparse, mostly uninformative sensor readings
• –The released Tiny encoder has just 5.46M parameters, making high-frequency robot inference practical
• –Cross-sensor transfer is the real test: useful tactile features should survive new hands, layouts, and manipulation tasks