Predictive Motion Planning for Safe and Efficient Autonomous Driving

Abstract

The advent of Autonomous Vehicles (AVs) has the potential to revolutionize transportation systems, promising significant improvements in safety, efficiency, and passenger comforts. Safety, the cornerstone of AVs, demands that these vehicles complete their routes without any collisions. This is a challenging feat, given the nonstationary nature of driving environments, where changes are often dictated by other traffic participants. AVs are also expected to handle distribution shift problems, adapting seamlessly to varying weather conditions, road types, and environments, be it urban, rural, or highway. Interactions with humans on the road, who often exhibit unpredictable behaviors, add another layer of complexity. Furthermore, AVs must learn to tackle long-tail, less represented scenarios that are rare but critical. However, safety is not the sole determinant of the widespread acceptance of AVs. Passenger comfort is equally important, making it imperative for Avs to ensure smooth and pleasant rides. One way to address these challenges is to predict the trajectories of surrounding traffic participants, thereby understanding how the driving context will evolve in the future. Yet, predictions can be with uncertainties. Therefore, the motion planning module should be designed to model these prediction uncertainties, ensuring safe planning.

This thesis introduces Deep Spatio-Temporal Context-Aware decision Network (DST-CAN), a novel framework for predictive manoeuvre decisions in Autonomous Vehicles (AVs). DST-CAN predicts future trajectories of surrounding vehicles, generating a spatio-temporal context-aware probability occupancy map. This map aids in making safe and efficient manoeuvre decisions. The performance of DST-CAN, evaluated using real-world datasets, shows superior performance over the state-of-the-art models. However, imitation learning method can suffer from distribution shift problem. On the other hand, traditional rule-based methods often fall short in complex real-world driving scenarios. Hence, this thesis introduces a new Predictive Planning Policy Network (P3Net) that uses Deep Reinforcement Learning for maneuver planning in AVs. P3Net addresses these issues by employing a predictive model and a Reinforcement Learning agent for decision-making. The RL agent uses spatio-temporal context to decide on maneuvers. Performance tests using real-world datasets show that P3Net continuously improves driving experiences, balancing safety and comfort, and outperforms recent imitation learning-based models in complex scenarios.

Also, modeling interactions between different traffic participants is crucial for the development of safe and efficient autonomous driving systems. It allows for a more accurate prediction of the future states of the traffic environment, enabling the autonomous vehicle to make informed decisions. For this reason, a new method is proposed in this thesis. A Graph-based Prediction and Planning Policy Network (GP3Net) for non-stationary environments has been proposed in the thesis. Evaluations on standard benchmarking scenarios show that GP3Net outperforms previous models, particularly in different towns and under new weather conditions. Further, modeling attention to surrounding traffic participants is a critical aspect of safe motion planning in autonomous driving. It involves the AV’s ability to focus on relevant traffic participants and scenarios, effectively filtering out less important information. To tackle this, a novel framework, Deep Attention Driven Reinforcement Learning (DADRL) is proposed. DADRL dynamically weighs the importance of surrounding vehicles in the AV’s decision-making process. The model, trained using the Soft-Actor Critic (SAC) algorithm, outperforms recent methods in different benchmarking scenarios. All these findings demonstrate the importance of Predictive Motion Planning (PMP) algorithms for autonomous driving for future improvements and adoption of this technology.