ETH Zurich Teaches a Robot Hand to Walk on Its Own Fingertips
ETH Zurich researchers taught a commercial robotic hand to walk on its fingertips, then use those same fingers to press keys and push objects.
Takeaways
- ETH Zurich trained a commercial anthropomorphic hand to walk on its fingertips using onboard compute.
- Hardware is a 20-joint hand plus battery and Raspberry Pi Zero 2 W, fully untethered.
- Reinforcement learning in a hardware-calibrated simulator, transferred through the existing position controller.
- A footprint reward outperforms quadruped-style rewards and gets all five fingers participating.
- Demonstrated crawling, steering, fall recovery, blind keyboard pressing, and vision-guided cube pushing.
- Paper on arXiv; no public code release announced yet.
Robotic hands usually rely on arms to carry them within reach of an object. An ETH Zurich team removed that dependency in Fingers as Legs, a research project that trains a commercially available anthropomorphic hand to crawl on its fingertips. After reaching a target, the hand reuses the same digits to press keys or push objects.
| Component | Implementation |
|---|---|
| Actuation | Twenty joints, with four joints per finger |
| Compute | Raspberry Pi Zero 2 W running the control policy onboard |
| Power | Battery mounted with the hand |
| Policy input | Recent movement, current hand state, and task goal |
| Policy output | Desired joint states sent through the existing position-control interface |
| External equipment | No tether or external computer for locomotion; an overhead camera supplies feedback for cube pushing |
Five fingers, unstable footing
Five unequal fingers create contact geometry unlike that of a conventional quadruped. The thumb is offset, each digit has a different reach, and joints designed for precise grasping must support the device’s full weight. During locomotion, the controller must coordinate which fingertips carry the body and which move forward. During a keypress, four contacts stabilize the chassis while a free finger moves toward the key. Those shared joints need enough stiffness to lift the hand and enough compliance to make controlled contact.