We are used to performing many everyday tasks with two hands: for example, to open a bottle, we hold it with one hand and twist the cap with the other. Researchers at the University of California, San Diego have developed EgoThumb - a robotic thumb that attaches to the arm and can take on part of the work, allowing some tasks to be handled with one hand. The paper was published on arXiv as a preprint, that is, before peer review.
The system includes a miniature wide-angle camera at the palm. It continuously captures the hand, the robotic thumb itself and surrounding objects. These images are analyzed by an artificial intelligence (AI) algorithm trained to plan the thumb's movements. AI is a program that learns from examples and then makes decisions.
The robotic thumb has three joints, each driven by a pair of tendons - one bends, the other extends. The motors are mounted on the forearm so as not to weigh down the hand. The scientists wanted the thumb to decide on its own when to grasp or release an object, relying on camera images and information about joint positions.
The AI model was trained by imitation: a human controlled the robotic thumb through a foot interface, creating training examples. After training, the model predicts thumb movements directly from camera images and current joint angles, without explicit user commands. In the prototype, the AI runs on an external computer connected to the device.
In laboratory experiments, three volunteers performed four tasks: lift a cube, unscrew a bottle cap, stir tea and move a ball from one place to another. In all cases, the extra thumb helped them manage with one hand. For example, when the user reaches for the bottle, EgoThumb automatically grasps and stabilizes it, while the user's fingers unscrew the cap.
The researchers also tested operation with different people and under different conditions. Two new participants and the original demonstrator showed comparable results when objects were in the same zones as during training. The average success rate across the four tasks with 10 attempts per task for each user was 90% for the demonstrator, 92.5% and 87.5% for the two new participants. In a separate test with tea stirring under two previously unseen conditions, the average success rate of the three users was approximately 67% and 73%, respectively. When the zones where objects were located or were supposed to end up were expanded, results on some tasks declined.
The authors emphasize that this is an early prototype, not a finished product. The EgoThumb hardware designs and software are open, and other developers can access them on the project website. The plans are to make the thumb more compact and convenient by improving camera integration, and also to add tactile sensors so that the robotic thumb can sense contact with objects and adjust its grasp.