Teaching artificial intelligence using the method of trial and error when the computer is “fed” recordings of the huge amount already wagered parties on the basis of which he hones his skill, proved that the machine is able to outperform humans in such classic Board games like chess and puzzle game. Of the most recent examples can be considered as the recent landslide victory of a computer over a person in a strategic computer game StarCraft II, where the machine was trained on the same principle. However, another group of experts in the field of AI has shown that this teaching method can be used for more practical tasks, for example, for training of robotic prostheses.
The method of machine reinforcement learning (reinforcement learning), in which the test system learns by interacting with some environment, showed promising results during a small experiment with a couple of volunteers – one fully healthy person and one with an amputated above-knee leg.
When using traditional methods the techniques usually takes several hours in order to properly configure the robotic prosthesis, manually adjusting each artificial joint and adjust it for a certain style of walking are accustomed to people. Experiment experts from the University of North Carolina showed that a machine learning technique reinforcement allows to do it much faster – 10 minutes after the fully automatic settings a person can go smoothly.
“To the real application of this technology is still very far. We just showed that this is possible. We admired”, says Helen Huang, Professor of bioengineering at the University of North Carolina.
Huang and his colleagues published their findings in the journal IEEE Transactions on Cybernetics. The results of their work can be an important first step towards the automation of typical processes for manual adjustment of robotic arms, which usually takes quite a long time and demands from patients ‘ visits to specialists whenever dentures require adjustment. In the future, all these settings people will be able to perform at home by yourself, without the help of technicians.
The very setting of the robotic prosthesis is a complex process of adjusting various parameters defining the levels of interaction between the limb and the prosthesis required for the performance of certain tasks. For example, some parameters determine the level of hardness of a robotic knee articulation or the tolerances allowed when rolling the foot back and forth. In the discussed case the knee robotic prosthesis required settings 12 different parameters. In the standardized approach, the end result usually obtained is far from ideal, but nevertheless was quite suitable in order that man might stand on the prosthesis and to make simple movements.
Training the robotic limb is a very complex process of co – adaptations. Prosthesis literally have to learn to work in tandem with the human brain, the Manager of the mutual adaptation of organs in a whole organism. At the same time, learning to walk need not only the prosthesis, but the man. Typically, the first results look quite awkward near the examples with skis or ice skates, which one first stood up.
“Our body can be quite strange to react to foreign objects that mimic its sequel. In a sense, our computer is a machine learning algorithm with reinforcement learning interaction with the human body,” says co-author in a published study, Jennie si, Professor of electronic, computer and energy engineering from the State University of Arizona.
The task of learning robotic prostheses is complicated by the very limited set available to the learning algorithm data. For example, to train their algorithms AlphaZero and AlphaStar for playing chess, go and StarCraft II company DeepMind used the record of millions of already played games of these games. In turn, people with an amputated limb to collect the necessary data to train the algorithm will not be able to walk very long amount of time. For example, those who visited the laboratory Huang was able to walk without stopping only 15-20 minutes, after which they needed a little rest.
But this is not all the difficulties and limitations, not allowing to cover the whole range of educational information, the researchers note. For example, even between si and Huang before the beginning of the project has arisen some dispute about whether or not to allow volunteers involved in the experiment, falling below the algorithm is able to learn this information. In the end, this idea decided to refuse, be attributed to the security volunteers.
And yet, even despite all these difficulties the first results impressed. The researchers trained the algorithm to identify particular patterns in the data collected by sensors installed in the robot knee. This, in turn, has allowed to establish the threshold functionality of a robotic prosthesis, allowing to avoid unwanted situations that could cause you to fall. In the end, the algorithm has learned to rely on a specific pattern, thereby to achieve stability, smoothness and more natural in the movement of robotic limbs.
Automated learning approach of robotic arms is still very far away from mass application. Now scientists want to teach the algorithm to a smooth control of the prosthesis when standing, lifting (e.g., chair) and the descent (e.g., stairs). In addition, the task is to make the system more Autonomous, allowing for training and adjustment of the prostheses, not only in laboratory conditions.
One of the most difficult and at the same time most important tasks, according to researchers, is to develop a method of “communication” algorithm and human, to enable them to tell him which of the selected settings of the prosthesis is most convenient. Early attempts to solve this problem using standard buttons and other simple methods of information input were ineffective. Perhaps partly because this version of the interface “computer-machine” fails to convey the completeness of perception of the coordination of man.
“This method did not work because we do not fully understand all the features of the human body. First, you need to fill in some gaps in fundamental knowledge about psychology and physiology,” concludes Huang.
Prospects of development of robotic prostheses based on artificial intelligence can be discussed in our Telegram chat.