ETH Zurich robot hand walks around on its own fingers
No wheels, no legs, no separate drive mechanism, and the power and compute ride onboard.
A robotic hand built at ETH Zurich holds up its own weight and crawls across a surface using the same fingers it grasps with. Lukas Ziegler wrote up the work in a thread, and his framing is the fun part. No wheels, no legs bolted underneath, no separate drive mechanism.
The same fingers that grasp things also carry the hand across a surface and hold up its weight.
A hand that walks to a task and then performs the task is a mobile manipulator with no locomotion hardware in it.
Why a hand is not a quadruped
Most legged locomotion reward functions (the scoring rules that tell a policy in training what counts as good walking) assume four identical legs spaced evenly around a body. A hand does not look like that. The fingers are different lengths, they sit unevenly around a palm, and the thumb matches none of them.
So the team wrote a reward formulation for that unequal geometry. Per the thread, in simulation it moves faster than tuned rewards borrowed from quadruped work. That is a simulation result against baseline reward functions, not a measured speed on hardware.
What it actually does
Ziegler lists crawling, steering, and recovery after being knocked over. While holding itself up, the hand executes a sequence of keyboard commands with no vision at all. It also pushes an object to target positions using overhead visual feedback. The simulator was calibrated against measurements taken from the physical hardware.
Two details make this more than a stunt. The team kept the hand's existing finger design and its position controller, so this is not a bespoke machine built for the demo. And power and computation sit onboard, so it runs untethered.
Shown versus claimed
This is a paper demo, not a product. The thread credits @amrhkzp among the authors and links the paper. Nothing in the source says how fast the hand moves in the real world, how long it runs on its onboard power, or what the hand is meant to be used for.
The idea underneath is the interesting bit. A hand that walks to a task and then performs the task is a mobile manipulator with no locomotion hardware in it at all. Every other route to that capability involves adding a base, or legs, or a drive train, and then carrying the weight and cost of the thing you added. Here the manipulator and the drive are the same set of fingers. Whether that scales past a lab surface to real floors, clutter and payloads is not something anybody has shown yet.