.jpg)
Researchers at Analog Devices have trained a stationary five-finger robot hand to trace an unseen Ethernet cable using pressure maps from its fingertips and joint positions, without a camera or an explicit model of the cable. In matched 60-hertz trials, Touch2Trace raised the share of runs that reached 10 centimetres from 0% with joint positions alone to 93% with dense tactile input; mean distance rose from 0.2 to 20.1 centimetres. The work has been accepted to CoRL 2026.
Tracing means repeatedly pinching and curling the thumb and index finger so the cable slides through the hand without escaping. Flexible objects change shape under contact, and the pressure needed to retain them shifts from moment to moment. The team collected 12 training demonstrations, 10.1 minutes in total, on one USB cable arranged as a ring. Behavior cloning, a form of imitation learning, then trained the policy to reproduce the operator's joint commands.
At each control step, custom 32-by-32 pressure arrays on the thumb and index finger send two tactile images to a self-supervised encoder that had learned contact patterns from about two million simulated frames and 30,000 real frames. The compressed touch features are combined with eight joint positions and the preceding 15 observations, covering 250 milliseconds. A small transformer converts that sequence into eight new joint targets. The 4.1-million-parameter pipeline took 3.4 milliseconds on an Nvidia RTX 4500 Ada GPU and ran at 60 hertz.
The main comparison held the policy architecture, training data and 60-hertz control rate constant. Each condition ran 30 trials on an Ethernet cable in a straight routing that the policy had not seen during training; a rotary encoder measured distance independently of the policy's sensors. Dense tactile maps reached the 10-centimetre mark in 93% of trials. A 16-feature summary of pressure reached 30%, while joint positions alone reached 0%. The matching mean distances were 20.1, 7.2 and 0.2 centimetres.
| Policy input | Trials reaching 10 cm |
|---|---|
| Joint positions only | 0% |
| Compact tactile summary | 30% |
| Dense tactile maps | 93% |
The ablations show where that advantage came from. Freezing the pretrained tactile encoder produced 20.1 centimetres, compared with 4.7 after fine-tuning and 0.2 from random initialization. Reducing control from 60 to 30 hertz cut mean distance to 5.7 centimetres; 15 hertz managed 0.5. Downsampling the pressure image to 8 by 8 produced 6.2 centimetres. These tests suggest that the robot needed learned spatial detail, a quarter-second history and rapid updates together.
The scope remains narrow. The hand was fixed without an arm, only eight of 20 joints were controlled, the sensor measured normal pressure but not shear, and training used one cable in one routing. Transfer was tested on three other cables and two routings, not on an assembly line or a changing workspace. A practical system still needs wrist motion, re-grasping and tests across more materials and failures. The 93% result therefore establishes dense, high-rate touch on this testbed; it does not establish general cable handling.
Sources
- Matteo Grimaldi and colleagues, “Touch2Trace: Tactile-Driven Imitation Learning for Dexterous Cable Tracing,” arXiv, September 14, 2026. Abstract, authorship, submission date, acceptance statement and headline result.
- Full paper. Hardware, data collection, policy design, matched trials, transfer tests, ablations, latency and limitations. The comparison table was created by Global Tech News from Tables 1 and 5. Lead photograph: Jérémy Barande / Ecole polytechnique Université Paris-Saclay, CC BY-SA 2.0; the image is illustrative and does not show the reported experiment.
- Tesollo DG-5F-M product page. Manufacturer description of the 20-degree-of-freedom robotic hand used as the research platform.