Researchers at ETH Zurich's Soft Robotics Lab have built a robotic hand that can walk on its own fingertips, and if that description alone is not enough to make you uneasy, the footage will finish the job. The project is being compared, quite accurately, to Thing from The Addams Family.
The hand is built on a standard commercial anthropomorphic model with five fingers and 20 actuated joints, four per finger. Rather than redesigning the hand, the researchers kept the off-the-shelf fingers as-is and taught the hand to move using reinforcement learning in a simulator. A battery, onboard computer via Raspberry Pi, and motion sensors were mounted to the palm to make it fully self-contained. The whole thing weighs 818 grams.
Teaching it to walk was a big challenge. Unlike a standard robot with symmetrical legs, the fingers are different lengths, the thumb opposes the others, and lifting any single finger to take a step immediately destabilizes the structure.
The neural network trained in NVIDIA's Isaac Lab simulation platform had to figure out independently when to lift each digit, rewarding stable posture and forward progress until the hand learned its own gait. In testing, it crawled across 14 different indoor and outdoor surfaces, including asphalt, tile, grass, gravel, and damaged stone, and when tipped over, it recovered and righted itself in 21 out of 25 attempts.
The same fingers it uses to walk switch back to functioning as fingers. In demonstrations, it approached a keyboard and pressed the correct keys with 90% accuracy, completing a level of the puzzle game Sokoban. It also pushed a small cube to a designated target position with millimeter precision.
Story Q&A
Powered by TweakBot
From TweakTown's coverage of this story.
How does the hand coordinate its unequal fingers and opposing thumb to stay balanced while walking?
How does the hand’s walking performance vary across the 14 surfaces tested?
What conditions affected the hand’s ability to right itself after being tipped over?
How could the walking hand be deployed to reach and operate controls in confined spaces?
Want something else? Ask TweakBot.
The practical application the researchers have in mind is placing the hand near a confined space a full robot cannot reach, letting it crawl inside, interact with an object or control, and return. Whether that vision plays out is a question for future development. For now, the main thing (no pun intended) it has proven is that a disembodied robotic hand walking toward you is exactly as unsettling as it sounds.
