Robot tactile model hits 65 percent success, nearly double best VLA

Vision-language-action models have made real progress on everyday jobs like folding laundry and tidying rooms, but they still stumble on tasks needing fine hand control — plugging in a USB cable or turning a key — because they ignore the tactile feedback humans rely on. The Berkeley work attacks that gap by pretraining on existing data first, then adding touch through a specialist submodel trained on 100 hours of demonstrations across more than 200 household objects, with a tactile expert running four times faster than the action expert so grip corrections arrive in time.

The field is also wrestling with data fragmentation. Chengbo Yuan at Tsinghua aggregated more than 3,000 hours from public datasets spanning 21 sensor types, converting each sensor's output into a shared format mapped onto a template of a human hand — an approach inspired by the Open X-Embodiment collaboration. That hardware-agnostic model beat a baseline even on hardware it had never seen, which Yuan attributes to a kind of common sense of tactile knowledge. Meanwhile USC's model infers touch from visual data alone, giving sensorless robots a rudimentary sense of touch.

Not everyone is convinced scale alone is the answer. Long Cheng of the Chinese Academy of Sciences argues tactile signals are sparse and intermittent, so models learn to ignore them; his IROS 2026 model predicts what a robot will feel from vision and amplifies only surprising mismatches, averaging 62.8% success against 28.2% without touch. The outcome now hinges on whether bigger, more diverse datasets or smarter algorithms — or both — prove decisive for truly dexterous tasks.