ETH Zurich preprint shows an off-the-shelf robot hand crawling on its fingers

Researchers add an 80 g onboard module to a commercial hand and train it to crawl, steer, right itself and press keys.

Researchers at ETH Zurich have trained a commercial anthropomorphic hand to use its fingers as legs. The work is a preprint posted to arXiv on 15 September and has not been peer reviewed; IEEE Spectrum reported it on 8 October.

The authors fit a 738 g Wuji right hand with an 80 g module carrying a battery, a motion sensor and a small computer, so it runs untethered. Separate policies, trained in a simulator calibrated against the hardware, handle crawling, steering, fall recovery and interaction.

The authors report crawling on 14 surfaces and righting the hand in 28 of 32 simulated falls. While supporting its own weight, the hand completed a Sokoban puzzle level with 12 keyboard presses, without vision, and pushed an object to targets guided by an overhead camera.

Sources

• arXiv preprint (Kazemipour, Zheng and Katzschmann, arXiv:2609.17172)
• IEEE Spectrum (secondary)

Written by the RankedRobot news desk (AI-assisted, human-reviewed) from the sources above; checked 2026-10-09. Claims are attributed to whoever makes them. Spotted an error? corrections@rankedrobot.com