A computer algorithm has significantly accelerated the adaptation of people with an amputated leg to a robotic prosthesis. As a result, the time for starting independent walking was reduced from a few hours to minutes. A description of the development was published in the IEEE Transactions on Cybernetics.
One of the most promising options for modern artificial intelligence is reinforcement learning. Within the framework of this approach, the computer program automatically tries to solve the problem, while receiving "rewards" for good results. The AlphaZero program is based on this principle, which has been able to win the best human athletes in the board game Go, and also performs impressively in other games such as chess, Dota 2 and Starcraft II.
In the new work, the researchers took a similar approach to developing software for a robotic leg prosthesis. The result was tested on one patient with an amputated limb just above the knee. Usually, such people need many hours of training to move without assistance, as well as lengthy manual adjustments of the prosthesis by specialists to match the individual characteristics of walking. The new algorithm allowed the disabled person to walk on their own after 10 minutes.
Setting up the prosthesis is complex and includes 12 parameters that control properties such as knee stiffness and range of motion when swinging the leg. The algorithm was trained on a fairly small set of data collected by special sensors on prostheses of other people with amputated legs. The authors note that despite the impressive results, the algorithm is still far from large-scale application, since at the moment it is designed only for walking on a straight surface. In the future, the researchers are going to add new movements, primarily climbing and descending stairs, as well as test the development on more people in need.