ETH Zurich Researchers Train 5-Fingered Robotic Hand to Walk on Fingertips
Researchers at ETH Zurich trained a 5-fingered, 20-joint robotic hand to walk autonomously on its fingertips while preserving fine manipulation skills. Powered by deep reinforcement learning in the NVIDIA Isaac Lab simulation framework, the untethered 818-gram device successfully traversed 14 different indoor and outdoor surfaces. This work demonstrates that off-the-shelf robotic hands can learn dual capabilities—locomotion and delicate manipulation—without physical hardware redesign. It paves the way for detached robotic appendages that can autonomously crawl into tight, hazardous spaces to operate machinery or inspect equipment. The autonomous hand integrates an onboard Raspberry Pi computer, batteries, and motion sensors, allowing it to self-right in 21 out of 25 push-over tests. In addition to fingertip walking, the hand achieved 90% typing accuracy on a keyboard and demonstrated millimeter-level precision when pushing blocks in a Sokoban puzzle game.
## BACKGROUND
Sim-to-real transfer is a key robotics approach where AI neural networks are trained inside accelerated physics simulations before being deployed onto physical robots. Frameworks like NVIDIA Isaac Lab utilize GPU acceleration to simulate complex dynamics, enabling robots to master difficult balance and gait control without risking physical hardware damage during trial and error.