ETH Zurich Develops Autonomous Robotic Hand Capable of Fingertip Locomotion and Manipulation
A groundbreaking advancement in robotics has emerged from ETH Zurich's Soft Robotics Lab: a self-sufficient, five-fingered robotic hand capable of navigating diverse terrains, maintaining balance, recovering from falls, and interacting with its environment, all by utilizing its digits. This innovative design transforms the conventional robotic hand into a mobile entity, where its fingers perform dual functions of support, movement, and object interaction. The system leverages sophisticated reinforcement learning techniques, trained in simulated environments, to achieve complex motor policies that are then applied to the physical device. This approach signifies a substantial leap in untethered robotic manipulation, moving beyond traditional locomotion methods to integrate mobility directly into the end effector.
The prototype, weighing 818 grams, is built upon a standard anthropomorphic hand featuring five fingers, each equipped with four actuated joints, totaling 20 joints. Researchers Amirhossein Kazemipour, Hehui Zheng, and Robert K. Katzschmann integrated a compact onboard system, including a power source, sensors, and processing hardware, making the robot fully self-contained. This untethered design eliminates the need for an external arm to transport the hand, marking a significant step towards practical mobile manipulation. The team drew inspiration from its resemblance to a character from 'The Addams Family,' highlighting its unique and somewhat eerie independent movement capabilities.
The process of enabling a hand to 'walk' is inherently complex, differing greatly from the mechanics of a typical quadrupedal robot. The distinct geometries and ranges of motion of each finger, coupled with the natural angle of the palm, present unique challenges. Furthermore, when one finger is lifted for a step or to engage with an object, the remaining fingers must continually adjust to maintain the robot's overall balance. The research team overcame these challenges by employing reinforcement learning. This method involved meticulously training locomotion strategies within a simulated environment before transferring these learned behaviors to the physical robotic hand, allowing it to adapt to its asymmetric structure and dynamic balancing requirements.
The efficacy of this robotic hand extends far beyond simple laboratory demonstrations. The device successfully navigated 14 different indoor and outdoor surfaces without external power, including carpet, tile, metal grates, asphalt, grass, gravel, and weathered stone. Its resilience was further demonstrated by its ability to self-right after being deliberately placed on its side, succeeding in 21 out of 25 attempts. Moreover, the fingers seamlessly transitioned between locomotion and manipulation. In one notable experiment, the robot sustained its balance while accurately pressing keyboard arrow keys, achieving 29 correct inputs out of 32 attempts. It subsequently applied these newfound keyboard skills to successfully complete levels in the puzzle game Sokoban, showcasing its impressive versatility and fine motor control.
Further experiments explored the hand's capacity to approach and propel small objects towards a designated target, highlighting the integrated functionality of locomotion and manipulation within a single system. This dual capability forms the core of the research. Traditional robotic hands typically operate as extensions of a larger arm, often requiring a mobile platform for movement across space. By endowing the hand itself with independent locomotion, the need for some external hardware is reduced, potentially enabling larger robots to deploy the hand into confined or otherwise inaccessible spaces. The researchers envision future applications where a larger robot positions the hand near a narrow opening, allowing it to autonomously crawl to a control panel or object, execute a required task, and then return for retrieval. While further refinements in onboard perception and expanded movement repertoires are necessary, this experiment fundamentally redefines a common robotic component. Instead of designing additional mechanisms to transport the hand, ETH Zurich has pioneered a method where the fingers themselves facilitate its independent movement.