A research team has built the first unified multimodal model designed specifically for touch.
UniTac treats tactile perception as a transition from non-contact to contact — the moment a sensor physically meets an object. The model uses a dual-level representation that encodes both the sensor doing the touching and the object being touched. That distinction matters because the same object feels different depending on which sensor reads it. UniTac handles two understanding tasks — describing object properties and identifying which sensor produced a given signal — and generates realistic tactile outputs using a two-stage training process that reconstructs contact and then aligns it to sensor-specific priors.
Most multimodal research stacks vision, language, and audio while treating touch as a footnote. Robotics and prosthetics research has long argued that reliable tactile data is what separates a robot that can fold laundry from one that can only pick up a block. UniTac's cross-sensor design means a single model could generalize across the fragmented hardware landscape, where each sensor manufacturer ships its own data format and calibration assumptions.
The paper claims state-of-the-art performance on tactile benchmarks, trained on large-scale multi-sensor datasets — though "state-of-the-art" in a field this young is a low bar to clear.