Robots are still largely numb to the unknown.
Researchers have released RCT (Robotic Contact Tactile), an open-source dataset built from full robot presses on 122 industrial reference materials across seven categories. Three DIGIT sensors captured 29,279 tactile frames at multiple contact positions, and each press is stored as a contact sequence rather than isolated snapshots. That design choice turns out to matter a lot: when the researchers prevented training and test data from sharing contact sequences, tactile-to-text retrieval accuracy dropped by 17.7 percentage points. Hold out entire materials at training time and performance falls further, landing at a Recall@1 of just 25.1% on unseen materials. That is not a rounding error — it is a blinking warning light.
The dataset exists because prior benchmarks were quietly cheating. The researchers show that a widely used public split (TVL/HCT) has test contact sequences fully present in training, and that raw-pixel nearest-neighbor search recovers the correct sequence 98.3% of the time — meaning a model can ace that benchmark without learning anything useful about touch. RCT is designed to close that loophole with evaluation splits that are actually hard. The finding reframes tactile generalization not as a solved problem but as an open one that the field has been underestimating.
Robotics labs have poured effort into vision and language grounding, but the sense of touch has lagged behind both in data and in rigorous evaluation. RCT does not hand anyone a solution — a 25% recall rate on novel materials is a long way from useful — but it at least gives researchers a benchmark they cannot game with a nearest-neighbor lookup.