PhD Defence | Learning Better
VU Amsterdam PhD candidate, Gongjin Lan will defend his PhD thesis “Learning Better – From Baby to Better” on Wednesday 16th December.
Gongjin Lan’s thesis focuses on the Infancy phase of the Triangle of Life (ToL) for learning effective locomotion skills in modular robots with evolvable morphologies. Lan completed his research under the supervision of Guszti Eiben, Evert Haasdijk and Jakub Tomczak (all from VU Amsterdam).
In this thesis, the overall long term vision behind his line of research is the concept of the Evolution of Things, to have a programmable evolutionary system that works with physical artefacts, i.e., robots. The Triangle of Life describes a generic architecture of robot systems, where both morphologies and controllers undergo evolution.
Locomotion is a fundamental task for designing robots that carry out other useful tasks. First, they studied how to learn directed locomotion, then they investigated targeted locomotion. These algorithmic studies are conducted in simulated and physical modular robots. To this end, they investigated learning algorithms in terms of data and time efficiency. They also investigated real-time robot vision to identify and follow the targets. Finally, they investigated multi-robot intelligence to learn behavioural strategies for a collective task.
This thesis contains several contributions. Firstly, a method for directed locomotion on modular robots in simulation and the experimental validation in the real world. Secondly, a time-efficient black-box optimizer, the Bayesian-Evolutionary Algorithm, that can be used to learn adequate robot controllers. Thirdly, a method for targeted locomotion on modular robots to approach and follow targets. This approach contains two key ingredients, a new closed-loop controller architecture with sensory feedback and an internal frame of reference to distinguish left and right in any body shape. Fourthly, two methods of real-time robot vision running on a Raspberry Pi to recognize objects. Finally, to generalize the tasks for one robot to a group of robots, we investigated multi-robot intelligence by combining simulated and real-world evolution of behavioural strategies.
In summary, this thesis provided the methods to learn more about brains for the overall long term vision of the Evolution of Things concept. That is, for robot systems where both morphologies and controllers evolve in real time and real space. Such an evolutionary robot system can be used to generate adequate robots for challenging applications. The work reported in this thesis provides a contribution to this vision by demonstrating how ‘newborn’ robots can learn skills and tasks quickly after birth.