Contents I. Introduction
With an ever increasing trend in the use of handheld devices and a consequent enormous explosion in data generated, researchers have been trying hard to figure out varied techniques to learn from such data by aggregating the same in a centralized entity. The learning algorithm is then run on this central server and the knowledge gained is sent back to all connected participating devices. This process is not always viable considering the fact that a large amount of data needs to be uploaded to the server and then processed periodically [1]. One work-around is to use a technique termed Federated Learning (FL) [2] where, in lieu of data, the models generated in-situ or on-device are shared by all participating devices with the central server. The server, in turn, aggregates them using some pre-defined techniques [3] –[5] and sends the modified model back to the devices. Thus, every device has a local dataset using which, a model, most often an Artificial Neural Network (ANN), is trained and evolved. Each device periodically sends the trained weights of its respective ANN to the central server in the form of an update. At the server these weights are, for instance, averaged, and then sent back to all the devices. Most often these new set of (averaged) weights represent a part of the learning performed at each of the participating devices and hence contribute to the enhancement of learning within the network. Over several rounds of this process, the models at each of the devices saturate to fairly homogeneous ones. This centralized version of FL model suffers from inherent drawbacks such as a central point of failure, scalability, privacy issues, coupled with the requirement of large clients-to-server bandwidth [1]. In order to overcome these hurdles, decentralized versions of FL have been proposed [6] –[10]. FL has also made its niches in multi-robot scenarios [11], [12]. In such cases, each robot shares its learned model with others, thereby aiding in faster learning convergence. The robots could be either in the same environment or different ones. Robots, most often need to connect to a centralized server, a cloud or a controller, which in turn performs the task of aggregating the models received. Since robots could be mobile, their dynamically changing positions may tend to make or break connections with the central entity. For a group of mobile robots in the same environment or different environments, a decentralized approach or a hybrid of the centralized and decentralized approaches, could prove to be more beneficial. Research on FL in the area of robots, most often target specific or customised robotic scenarios, making it difficult for others to reuse the work.
