Aggregation Mechanisms in Federated Learning for Enhancing Robotic Visual Obstacle Avoidance

In the realms of robotics and the autonomous driving field, deep learning has been a game-changer for applications such as perception, obstacle avoidance, SLAM, etc. The success of these deep learning algorithms, however, relies on large-scale data sets for training, raising substantial privacy concerns for users unwilling to share raw data. Addressing this issue, our study explores federated learning in robotics to leverage its inherent privacy-protection attributes. Federated learning, a privacy-conscious method within the machine learning paradigm, has become increasingly prominent in recent years. More importantly, an essential aspect of our work is the assessment of diverse aggregation mechanisms within the federated learning framework, aiming at enhancing robotic functionality. We gather data and implement a federated learning model using various aggregation algorithms in visual obstacle avoidance contexts. The results demonstrate considerable variances in the performance of these algorithms, indicating that certain federated learning aggregation methods can exceed the performance of traditional centralized learning models. This study affirms the feasibility and advantages of federated learning in robotic applications, publicly available on github.