Architecture for real-time coordination of multiple autonomous mobile units

Interest on using teams of mobile robots has been growing, due to their potential to cooperate for diverse purposes, such as rescue, de-mining, surveillance or even games such as robotic soccer. These applications require a real-time middleware and wireless communication protocol that can support an efficient and timely fusion of the perception data from different robots as well as the development of coordinated behaviours. Coordinating several autonomous robots towards achieving a common goal is currently a topic of high interest, which can be found in many application domains. Despite these different application domains, the technical problem of building an infrastructure to support the integration of the distributed perception and subsequent coordinated action is similar. This problem becomes tougher with stronger system dynamics, e.g., when the robots move faster or interact with fast objects, leading to tighter real-time constraints. This thesis work addressed computing architectures and wireless communication protocols to support efficient information sharing and coordination strategies taking into account the real-time nature of robot activities. The thesis makes two main claims. Firstly, we claim that despite the use of a wireless communication protocol that includes arbitration mechanisms, the self-organization of the team communications in a dynamic round that also accounts for variable team membership, effectively reduces collisions within the team, independently of its current composition, significantly improving the quality of the communications. We will validate this claim in terms of packet losses and communication latency. We show how such self-organization of the communications can be achieved in an efficient way with the Reconfigurable and Adaptive TDMA protocol. Secondly, we claim that the development of distributed perception, cooperation and coordinated action for teams of mobile robots can be simplified by using a shared memory middleware that replicates in each cooperating robot all necessary remote data, the Real-Time Database (RTDB) middleware. These remote data copies, which are updated in the background by the selforganizing communications protocol, are extended with age information automatically computed by the middleware and are locally accessible through fast primitives. We validate our claim showing a parsimonious use of the communication medium, improved timing information with respect to the shared data and the simplicity of use and effectiveness of the proposed middleware shown in several use cases, reinforced with a reasonable impact in the Middle Size League of RoboCup.

Techniques for batch reinforcement learning in robotics

This thesis addresses the Batch Reinforcement Learning methods in Robotics. This sub-class of Reinforcement Learning has shown promising results and has been the focus of recent research. Three contributions are proposed that aim to extend the state-of-art methods allowing for a faster and more stable learning process, such as required for learning in Robotics. The Q-learning update-rule is widely applied, since it allows to learn without the presence of a model of the environment. However, this update-rule is transition-based and does not take advantage of the underlying episodic structure of collected batch of interactions. The Q-Batch update-rule is proposed in this thesis, to process experiencies along the trajectories collected in the interaction phase. This allows a faster propagation of obtained rewards and penalties, resulting in faster and more robust learning. Non-parametric function approximations are explored, such as Gaussian Processes. This type of approximators allows to encode prior knowledge about the latent function, in the form of kernels, providing a higher level of exibility and accuracy. The application of Gaussian Processes in Batch Reinforcement Learning presented a higher performance in learning tasks than other function approximations used in the literature. Lastly, in order to extract more information from the experiences collected by the agent, model-learning techniques are incorporated to learn the system dynamics. In this way, it is possible to augment the set of collected experiences with experiences generated through planning using the learned models. Experiments were carried out mainly in simulation, with some tests carried out in a physical robotic platform. The obtained results show that the proposed approaches are able to outperform the classical Fitted Q Iteration.