Fingers as Legs

A self-contained robotic hand that learns to crawl, recover, and manipulate using its own fingers.

An anthropomorphic hand becomes a compact mobile manipulator. The same fingers move the hand, support its weight, and interact with the world, with power and computation carried onboard.

Amirhossein Kazemipour, Hehui Zheng, and Robert Katzschmann · Soft Robotics Lab, ETH Zurich · 2026

Project website · Paper on arXiv · Watch the film

A robotic hand standing on its fingertips, with its battery and onboard computer above the palm
The self-contained platform weighs 818 g, including onboard power and computation.

Learning to move and interact

Lifting a finger to take a step or press a key changes how the hand supports itself. Our reinforcement learning approach accounts for its unequal fingers and trains in a simulator calibrated from hardware measurements. The learned policies run onboard, while retaining the hand’s original finger design and position controller.

On the real hand

  • Untethered crawling demonstrated on 14 indoor and outdoor surfaces under operator control.
  • Fall recovery, keyboard interaction, and object pushing while supporting its own weight.
  • Task-specific policies for each skill. Keyboard trials use manual initial alignment and no visual feedback; pushing uses an overhead camera.

The result is a hand that combines locomotion and manipulation without a separate walking mechanism. The research website includes demonstrations and interactive explanations of the learning approach.

Paper

Fingers as Legs: Learning Self-Supported Locomotion and Manipulation with an Anthropomorphic Hand. Amirhossein Kazemipour, Hehui Zheng, and Robert Katzschmann. arXiv:2609.17172, September 2026.