ETH Zurich Researchers Develop Robot Hand That Walks on Fingertips
Researchers at ETH Zurich have created a robot hand capable of walking independently on its fingertips. The hand learned locomotion through reinforcement learning and can perform other tasks.
Researchers at ETH Zurich (Swiss Federal Institute of Technology) have developed a robot hand that can independently walk on its fingertips. The development combines advanced robotics and machine learning, enabling the machine to navigate challenging environments.
The robot hand is a modification of a conventional humanoid robotic hand, featuring five fingers with four joints each. The device is equipped with a battery, a Raspberry Pi-based onboard computer, and motion sensors, allowing for fully autonomous operation. The entire unit weighs 818 grams.
Training the hand to walk involved reinforcement learning within the NVIDIA Isaac Lab simulation platform. The neural network had to learn when to lift and lower each finger to maintain balance and achieve forward movement. In tests, the hand successfully traversed 14 different indoor and outdoor surfaces, including asphalt, tiles, grass, and gravel. It also managed to right itself autonomously in 21 out of 25 instances when toppled.
Beyond locomotion, the robotic hand retains its ability to perform standard hand functions. It demonstrated accurate typing on a keyboard with 90% success and successfully played a level of the Sokoban puzzle game, moving small blocks to designated targets with millimeter precision.
Potential applications envisioned by the researchers include deploying the robotic hand into narrow spaces inaccessible to larger robots, allowing it to operate machinery or interact with objects before returning. Further research and development are needed to validate these practical applications.