Adding touch made a robot worse at zipping a bag. Translating the signals lifted its score from 15% to 59%

A robot given fingertip touch sensors became worse at closing a zipper. In a technical report submitted November 9, Ant Group's Robbyant team reported that a π0.5 policy controlling a Sharpa Wave hand scored 39% without touch, 15% with raw tactile readings, and 59% with LingBot-Touch features. Those scores award credit for individual task steps; they are not percentages of bags fully zipped.

Different tactile sensors produce different kinds of data, including deformation, magnetic readings and resistance changes. LingBot-Touch turns those signals into shared features that a robot's action model can use. Chinese technology outlet QbitAI covered the research, and the team has made its report, inference code and model weights available.

The team also tested a hand whose magnetic sensor was absent from pretraining. It kept the touch encoder frozen while training the action policies. Across six real-world XHand1 tasks, π0.5's average score rose from 48% with raw touch data to 74.17% with encoded features. That points toward carrying tactile experience from one hardware platform to another.

Zippers, nested cups and bottle caps all demand feedback from contact points a camera may not see. This preprint offers a way to supply it, but the results come from the team's controlled experiments. Action models still need demonstrations and training, and some tests did not improve. Reliable household manipulation remains a broader challenge.

Source times: arXiv technical report, November 10, 2026, 02:12 Beijing / November 9, 2026, 1:12 PM EST; QbitAI, November 10, 2026, 15:26:53 Beijing / November 10, 2026, 2:26:53 AM EST.