Robotic hand learns to walk on its own fingers

A humanoid robotic hand is learning to walk on its own fingers — and neither the finger design nor the stock position controller is changed. The work was posted to arXiv as 2609.17172 on September 15, 2026; the authors are Amirhossein Kazemipour, Hehui Zheng and Robert Katzschmann.

The core idea is reusing the fingers. The same fingers move the body, hold up its weight and interact with objects, while the hand has no separate mechanism for locomotion.

It is the fingers that play the role of legs here: bracing on them, the hand shifts its weight, crawls, steers and gets up after falls. No separate cart or wheels are needed for that.

Power and computation sit on the hand itself rather than on an external rig, so the platform is self-contained and works without a cable.

The skills were obtained through reinforcement learning. The simulator was calibrated against measurements of the real device to account for the fingers not being equal: they differ in length, and the algorithm itself decides which finger to reposition while the others hold the weight. In simulation, the hand moved faster with the reward proposed by the authors than with a tuned reward originally designed for quadruped robots.

On real hardware, task-specific policies deliver untethered crawling, steering and getting up after a fall. While bearing its own weight, the hand executes sequential keyboard commands without vision, and pushes an object toward a target using feedback from an overhead camera.

The authors sum up the result this way: what came out of it is a compact mobile manipulator that reuses its fingers both for locomotion and for interaction, doing without a separate drive mechanism.

Image: Giacomo Alessandroni CC BY-SA 4.0 · License

Translation editor: Салтанов А.