Physix FrontierPhysix Frontier · News Briefing Card (QbitAI · Jul 18, 2026)
UniTouch Releases UniTac for Cross-Sensor Tactile Prediction
KEY FACTS
- UniTouch, together with multiple universities, released the UniTac architecture, accepted to ECCV 2026.
- The model unifies multi-sensor data, enabling robots to predict tactile states before contact.
- UniTac-7B scored 66.51 on the PHYSICLEAR benchmark, outperforming existing models.
- The team trained it on over 400,000 segments of scattered public tactile video data.
KEY DATA
66.51PHYSICLEAR Score
0.836Average SSIM
19.93Average PSNR
400,000+ segmentsTraining Data
PHYSIX OBSERVATION
Touch is shifting from passive feedback to an active prior, a key leap for embodied intelligence deployment. UniTac solves the data-silo problem caused by heterogeneous sensors, bringing touch into the foundation-model loop. This "sense-before-touch" capability significantly reduces the risk of manipulating deformable objects, signaling that robots will move from coarse grasping to fine-grained physical interaction, clearing perceptual obstacles for unstructured scenarios such as home services.
Source: QbitAI report