Before a major Parisian event that will take place just before the summer, let’s shed light on robots using learning techniques directly inspired by psychology work on newborns.
The Futur en Seine event dedicated to innovation and startups will take place from June 8 to 10, 2017. An appetizer of this highly anticipated gathering took place on February 22, 2017 with Cap Digital, which brought together around twenty companies specializing in artificial intelligence.
Among these start-ups, there is Hoomano which works in collaboration with the Laboratoire d’InfoRmatique en Image et Systèmes d’information (LIRIS) on the development of learning software for machines. Their so-called “developmental” approach is in contrast to a more traditional “reinforcement” learning. While the latter lets the environment guide the machine, the developmental method allows the robot to explore its own environment driven by its search for satisfaction linked to its own success.
In case of failure, the machine can try by itself to find other solutions in order to circumvent possible obstacles such as a child who fails and acquires experience. With the reinforcement method, the environment must be known beforehand to inform the machine of the presence of an obstacle.
In reality, you would have to imagine a six-legged robot. In the event that the latter loses one of its limbs, the developmental approach will allow it to find a way to move with its remaining five legs, but the reinforcement approach will keep it immobile simply because its environment does not allow it to move. will indicate no solution.
These famous learning techniques developed by the Hoomano company and LIRIS are directly inspired by the work of Jean Piaget, a Swiss psychologist, biologist and epistemologist of the 1930s. Just as a newborn finds pleasure in sucking his thumb, the robot will experience pleasure (corresponding to an algorithmic sequence) in the success of its moves and movements.
The video below shot by Science et Avenir is very telling. A Pepper robot performs gestures by itself and Amélie Cordier, scientific director of the Hoomano company, validates « the right gesture » by pressing on its head. This is then a source of satisfaction for the machine which will reproduce the same gesture to have confirmation that its action is indeed that expected.
Sources: Hoomano – Sciences et Avenir