Dear DD2410 students,

My name is Matteo Iovino and I am a PhD student at RPL. My research is about autonomous learning of Behavior Trees to control industrial collaborative robots to perform a kitting task.

With kitting, th intended operation is retrieving specific tools and placing them in a kit-box to finally deliver to a human operator (or a fixed manipulator) to perform an assembly task. The environment the robot would operate in is an industrial one.

The purpose of the research is to exploit the modularity of the control structure the Behavior Tree is, in short life cycle production. In such a way, only the task-dependent sub trees will be re-programmed, leaving the more general parts of the structure untouched.

The learning part is also required because the final desired goal is a framework where the Behavior Tree is generated with some user defined inputs.

One way of autonomously learning a Behavior Tree is to use Genetic Programming. This approach starts from an initial pool of individuals and it assigns to each of them a value based on a fitness function. Thus, only the best individuals survive and can evolve. Then, through genetic operations (crossover and mutation), the algorithm populates the next generation of individuals. It proceeds creating new generations, lastly converging to the best individual. The idea of the research project is to apply Genetic Programming to learn the optimal solution for the task you are asked to solve. Your hand-codes Behavior Tree will be used as a dataset from where to pick the samples that populate the first pool of individuals. Moreover, your solutions will be used to evaluate the learned ones with respect to some metrics yet to be decided (e.g. complexity of the tree).

For further information you can find me at matteoi@kth.se