Êtes-vous un étudiant de l'EPFL à la recherche d'un projet de semestre?
Travaillez avec nous sur des projets en science des données et en visualisation, et déployez votre projet sous forme d'application sur Graph Search.
Robots outside of the fenced factories have to deal with continuously changing environment, this requires fast and flexible modes of control. Planning methods or complex learning models can find optimal paths in complex surroundings, but they are computationally expensive and hence not suitable for evaluation on-board. Dynamical systems (DS) as a mean to control robots allows to adapt the motion on the fly to outside disturbances and to continue the task without stopping. While theoretical collision avoidance for a velocity controlled robot can be guaranteed, the methods cannot ensure that an agent always reaches a goal. In this thesis, we investigate the problem of combining learned motion with reactive adaptation in a DS framework. We want to analyse asymptotic convergence to a desired motion by separating direction and magnitude of a desired motion (direction space). In the first year of the thesis, a closed-form approach to avoid a category of concave obstacles (star-shaped) has been developed. It applies to objects with non-smooth derivable surfaces, i.e. polygons with sharp edges. The algorithm guarantees that the robot will not penetrate the obstacles and reach a desired target. Additionally using an inverted description of the obstacles and the corresponding distance function allows to safely navigate inside a volume. These inverted obstacles can represent walls of a room or joint limits of a robot. These methods have been tested in simulation and on real robots in a laboratory environment. In future work, we plan to extend the current obstacle avoidance algorithm by focusing on the direction of the flow of the DS. We want to extend the algorithm to be able to handle uncertainties in environment prediction. Additional constraints will be observed to ensure a safe motion. A similar problem arises, when defining a desired velocity field for an agent, but controlling in force and torque. We hope to tackle these two problem by introducing directional repulsion in velocity and force, respectively. In a second part, existing learning methods should be extended and adapted to a description of a DS. We expect that separating direction and magnitude of the motion will allow to learn a wide range of motions while ensuring stability of the system. The last contribution should be a unifying frame work of learning motion and dynamic obstacle avoidance under velocity and force control, combining the work developed during this thesis. Continuous collaboration with colleagues and students is sought to accelerate and test real-world implementations. Since this requires the analysis of sensors, considering constraints and controllers of the robots as well as the extension of the algorithm to joint-space.
Maryam Kamgarpour, Orcun Karaca, Dániel Tihanyi