Less than two years ago, DARPA launched the Lifelong Learning Machines (L2M) program to create continuously learning robotic systems with elements of artificial intelligence. The L2M program was supposed to lead to the emergence of self-learning platforms that could adapt themselves to a new environment without prior programming or training. Simply put, robots had to learn from their mistakes, and not learn by pumping up sets of template data in a laboratory environment.
The L2M program involves 30 research groups with different amounts of funding. Just the other day, one of the groups from the University of Southern California showed convincing progress in creating self-learning robotic platforms, as reported in the March issue of Nature Machine Intelligence.
The research team from the university is led by Professor of Biomedical Engineering, Biokinesiology and Physical Therapy Francisco J. Valero-Cuevas. Based on the algorithm developed by the group, which is based on certain mechanisms of the functioning of living organisms, a sequence of artificial intelligence actions was created to teach the robot to move on four limbs. It is reported that artificial limbs in the form of imitation tendons, muscles and bones were able to learn to walk within five minutes after running the algorithm.
After the first launch, the process was unsystematic and chaotic, but then the AI began to quickly adapt to the realities and successfully started walking without prior programming. In the future, the created method of lifelong training of robots without preliminary ML training with data sets can be adapted for equipping civilian cars with autopilots and for military robotic vehicles. However, this technology has much more prospects and areas of use. The main thing is that the algorithm does not perceive a person as one of the obstacles in development and does not learn anything bad.
Source: 3dnews.ru