EgoThumb: a robotic sixth finger that works without being steered

How does EgoThumb work?

The build is simple. Three joints, each moved by a pair of tendons for bending and straightening. The motors sit on the forearm, which keeps the thumb light. A small wide-angle camera by the hand watches your fingers and the objects around them.

The camera feed goes to an AI model, along with the current joint positions. It was trained by imitation learning, meaning it learned from human demonstrations. During training an operator moved the thumb with a foot-operated interface. Afterwards the algorithm predicted the moves itself. Reach for a bottle and the thumb holds it while your fingers twist the cap off. Put the bottle down and it lets go.

Cambridge researchers built a Third Thumb earlier, steered with toe pressure. EgoThumb tries to remove the need for constant attention.

How well does it work?

Three people tried four tasks: lifting a cube, unscrewing a cap, stirring tea and moving a ball, ten attempts each. Success rates averaged 90%, 92.5% and 87.5%. Two participants who were not part of the model's training did about as well as the operator who supplied the demonstrations.

There is a catch. Those results come from places the algorithm knew from training. Move things to new surroundings and some tasks got worse: stirring tea in two new environments averaged roughly 67% and 73%. Three people is a small sample, the model runs on an external computer, and nobody has tested actual soldering yet.

Safety is the open question. A camera can miss an object hidden by the hand, and it cannot sense pressure or slipping, so a wrong call could mean a dropped or crushed item. The team plans to add touch sensors, make the device smaller and more comfortable, and integrate the camera better. Hardware designs and software are open source.