Principal Investigator: Juan Andrade; Joan Solà
Email: cetto@iri.upc.edu
Web: www.iri.upc.edu
Brief Theme Description:
This research theme will aim to equip a hybrid aerial-contact robot with agile movement capabilities through Reinforcement Learning (RL) techniques. We strive to move beyond the traditional model-based predictive control (MPC) approach —which computes optimal trajectories one at a time— and replace it with an RL framework that enables the robot to acquire a set of agile, hybrid motor skills (control policies) in advance.
The learning process begins in simulation, initially requiring a model, and subsequently transfers the learned control policies to the physical robot, which they must execute in real time. This approach allows for continued training and refinement on the real robot during the final stages, eventually eliminating the need for a model.
This project is part of a global effort to develop and implement complex, dynamic movements for multi-jointed robots—such as humanoids and quadrupeds—that use contacts for locomotion. However, the focus will be on hybrid aerial-contact robots, enabling them to perform sophisticated locomotion manipulation tasks that combine aerial and ground interactions.
Available Infrastructures: IRI Mobile Robotics Lab; Optitrack positioning system.
Possible Secondments: LAAS CNRS (France)
Keywords: Agile Robotics; Model Predictive Control; Reinforcement Learning; State Estimation.
