Distributed Robust Multi-Robot Learning using Particle Swarm Optimization

The goal of this project is the automatic design of high-performing robust controllers for mobile robots using exclusively on-board resources. In our evaluative approach, a population-based, on-line machine-learning technique automatically shapes the robots’ behaviors by direct interaction with the real environment. This constitutes an expensive optimization problem as the time needed to evaluate candidate solutions is substantially larger than that required by the metaheuristic operators in the algorithm.
In order to accomplish our goal, we must address two research questions. Firstly, we need to determine how to optimally allocate evaluation time for fast and robust adaptation, as evaluation time is the most critical resource in the process. Secondly, we need to define effective information sharing and cooperation mechanisms in our distributed system, as excessive sharing could lead to early stagnation, while limited sharing may slow down the convergence of the algorithm.

Team and Collaborators

• Ezequiel Di Mario
• Iñaki Navarro
• Guillaume Jornod
• Zeynab Talebpour

Sponsors and Research Period

National Center of Competence in Research Robotics (NCCR Robotics)  and SNSF grant

Videos

Video of 4 robots learning obstacle avoidance using distributed PSO

Publications