Imitation learning of elementary tasks with contacts

Proposé par : Serena Ivaldi (,


The team MAIA, and its probable sequel LARSEN participates to the European Project CoDyCo, whose aim is to advance the current control and learning techniques for whole-body motion of robots interacting with humans and environments, subjects to multiple contacts. CoDyCo is proposing methodologies for performing coordinated interaction tasks with complex systems, combining planning and compliance to deal with predictable and unpredictable events and contacts. The proposed algorithms are validated in real-world interaction scenarios with the iCub humanoid robot engaged in whole-body goal-directed tasks, such as balancing while reaching for a distant object, or leaning on a soft chair.

In this context, one of our objectives is to generalize and improve elementary tasks with contacts. We aim at generating a repertoire of elementary tasks to be useful in different scenarios, with a data-driven approach. We will record data on humans performing whole-body tasks with contacts, exploiting a motion capture system and a sensorized floor. Using both models and data, elementary tasks will be acquired by imitation learning. In this case, we will use a probabilistic description of the tasks, suitably formulated to be capable of taking contacts explicitly into account. Classical reinforcement learning techniques can be then used to optimize online the tasks, directly on the robot. This activity will involve a collaboration with the partners of CoDyCo in the Technical University of Darmstadt and Josef Stefan Institute.



C/C++, Linux, optimization, machine learning, basics of probability, robotics


Alain.Dutech AT