Learning Multi-Modal Self-Awareness Models Empowered by Active Inference for Autonomous Vehicles

For autonomous agents to coexist with the real world, it is essential to anticipate the dynamics and interactions in their surroundings. Autonomous agents can use models of the human brain to learn about responding to the actions of other participants in the environment and proactively coordinates with the dynamics. Modeling brain learning procedures is challenging for multiple reasons, such as stochasticity, multi-modality, and unobservant intents. A neglected problem has long been understanding and processing environmental perception data from the multisensorial information referring to the cognitive psychology level of the human brain process. The key to solving this problem is to construct a computing model with selective attention and self-learning ability for autonomous driving, which is supposed to possess the mechanism of memorizing, inferring, and experiential updating, enabling it to cope with the changes in an external world. Therefore, a practical self-driving approach should be open to more than just the traditional computing structure of perception, planning, decision-making, and control. It is necessary to explore a probabilistic framework that goes along with human brain attention, reasoning, learning, and decisionmaking mechanism concerning interactive behavior and build an intelligent system inspired by biological intelligence. This thesis presents a multi-modal self-awareness module for autonomous driving systems. The techniques proposed in this research are evaluated on their ability to model proper driving behavior in dynamic environments, which is vital in autonomous driving for both action planning and safe navigation. First, this thesis adapts generative incremental learning to the problem of imitation learning. It extends the imitation learning framework to work in the multi-agent setting where observations gathered from multiple agents are used to inform the training process of a learning agent, which tracks a dynamic target. Since driving has associated rules, the second part of this thesis introduces a method to provide optimal knowledge to the imitation learning agent through an active inference approach. Active inference is the selective information method gathering during prediction to increase a predictive machine learning model’s prediction performance. Finally, to address the inference complexity and solve the exploration-exploitation dilemma in unobserved environments, an exploring action-oriented model is introduced by pulling together imitation learning and active inference methods inspired by the brain learning procedure.