Robots learn

Uri Carton, a student currently completing his third degree in the Department of Industrial Engineering and Management at Ben-Gurion University develops artificial intelligence algorithms that are applied to robots

Figure 1. A robotic arm learning how to empty a bag containing non-conventional explosives (under laboratory conditions)

In the Department of Industrial Engineering and Management at the Ben-Gurion University of the Negev, under the guidance of Prof. Helman Stern and Prof. Yael Iden, Uri Karton, a student who is currently completing his third degree, is developing artificial intelligence algorithms applied to robots. Uri Karton's main research deals with the improvement of algorithms from the Reinforcement Learning (RL) group, a subfield of artificial intelligence. In RL an agent (for example, a mobile robot) functions in the world or in a certain environment. While performing actions in the environment, the agent receives reinforcements (Reinforcements) or punishments (Punishments), indicators indicating how satisfactory his conduct is. According to these indicators, the agent learns how to conduct himself in the environment, which situations it is better to avoid and which actions he prefers to perform.

Uri Carton developed an algorithm called , an algorithm that was applied to two robotic systems: (1) a robotic arm whose role is to grab a bag suspected of containing non-conventional explosives (such as SARS, anthrax and Ebola viruses) and empty it onto a test surface (Figure 1) and (2) a mobile robot for Navigation (Figure 2). The algorithm applied to a robotic system allows it to check if its learning level is insufficient or if it has encountered an unsolvable situation. If this scenario occurs, the system contacts the person for advice. The person offers a solution (strategy) when the assumption is that by virtue of being a person, he has an intelligent ability that exceeds that of the robotic system. The solution proposed by the person is received through an interface installed on a computer system remote from the robot when the communication between the person and the robot can be done via the Internet or via satellite communication.

The strategy offered by the person is taken into account by the learning functions of the robot and the more the number of times the person intervenes, the higher the learning level of the robot improves and its dependence on the person decreases. At a certain point, the robotic system's dependence on humans decreases to such an extent that it actually turns from a semi-autonomous system to a system with full autonomy, that is, the degree of human intervention provides a measure to judge the system's level of autonomy - the less intervention there is, the greater its level of autonomy and it becomes more independent and has More efficient intelligence.

The research carried out in the Department of Industry and Management can be applied to applications developed by NASA, for example, an attempt to disconnect and shake solar panels on the space station in December 2006. If a robotic arm could be taught to shake a solar panel on the ground of the Earth while dealing with different situations (such as: changing angles, different materials, different lighting conditions, etc.) while mixing the person in situations where the robotic arm would have encountered difficulties, the same knowledge that I had gained by the computer systems of the robotic arm while performing experiments on the ground, could be applied in space whether sharing The operation with the robotic arm would be carried out by an astronaut operator or whether the operator was on Earth.

Another example of the integration of this research into NASA systems is a system that deals with the execution of a navigation task by a mobile robot. the algorithm, can be applied to a robotic system whose job it is to survey the terrain of a distant planet in order to locate water sources. If the robot's task is to learn how to scan the space of a planet while avoiding limited obstacles and it encounters difficulties while performing its task, integrating a person into the learning process of the robot moving on the planet is worthwhile. Since planets are far from Earth and receiving the visual information and the location of the robot are not known at any given moment, the robot must function autonomously for at least part of its mission. The aforementioned lack of acknowledgment is usually due to limitations such as atmospheric concealment or poor weather conditions. If for part of the duration of the robotic task there is no problem of uncertainty, then it will be possible for the robot operator to assist him and thus bring him closer to the level of a perfect autonomous agent.

Figure 2. A mobile robot learning to perform a navigation task (under laboratory conditions)