Human-aware Autonomous Vehicle Navigation in Pedestrian-rich Unstructured Environments

Abstract

Autonomous Vehicles (AVs) have the potential to enhance transportation safety, improve efficiency, and elevate quality of life. Despite significant advancements in AV technology, operating these vehicles in dynamic, crowded environments that requires frequent interaction with other decision-making agents remains challenging. A key example is the interaction between AVs and pedestrians. While most research has focused on these interactions in structured road settings, the complexity and diversity of AV navigation among pedestrians in unstructured environments (e.g., shared spaces, airport terminals) have been less explored. In such pedestrian-rich environments, AVs must be human-aware, meeting people's expectations while ensuring both their safety and comfort. At the same time, navigating these spaces requires reasoning about mutual interactions and accounting for the uncertainty in pedestrian behaviour. This thesis introduces a novel approach to address these challenges, presenting an integrated prediction and planning framework for AV navigation among pedestrians in unstructured shared environments.

The thesis is structured into two main phases: a design requirement study and an algorithmic development phase. Given the novelty of this application, the first phase focused on understanding the perceptions and preferences of pedestrians regarding AV behaviour in common interactive scenarios within unstructured settings. Additionally, we examined the unique aspects of pedestrian behaviour in these environments, identifying common behaviours AVs must manage and gathering existing datasets that better represent pedestrian behaviour in such settings. This study highlights the importance of considering uncertainty in pedestrian behaviour, shaping the direction of the development phase.

In the algorithmic development phase, we propose a novel proactive, uncertainty-aware Deep Reinforcement Learning (DRL) decision-making algorithm. This algorithm efficiently accounts for complex interaction effects with multiple pedestrians while maintaining reasonable computational time. The navigation algorithm is made proactive and farsighted by integrating the DRL motion planner with a data-driven pedestrian trajectory predictor. Our novel prediction model is designed to forecast pedestrian trajectories in highly interactive shared environments. It uses a collision risk metric to identify key interacting agents and encodes their effects through a newly engineered interaction feature which guide the learning process. During training, we prevented overconfident predictions and improved estimates of prediction uncertainty using an augmented loss function that incorporates uncertainty awareness.

Unlike other DRL algorithms in this area, our model's DRL motion planner accounts for prediction uncertainty, integrating it into the reward function to encourage the AV to minimize collision probability with pedestrians over a prediction horizon. Additionally, the reward function design encourages socially aware behaviours, such as reducing speed during close encounters, respecting pedestrians' personal space, and adhering to social norms identified in our earlier design requirement study.

We trained our model in a simulation environment that contains realistic pedestrian trajectory behaviour in the presence of vehicles in shared spaces. The simulation results demonstrate that our uncertainty-aware DRL navigation framework outperforms state-of-the-art DRL crowd navigation and uncertainty-aware Model Predictive Control (MPC) models, both in terms of efficiency and social behaviour aspects.

Overall, this thesis contributes to the advancement of socially-aware crowd navigation algorithms beyond human-sized mobile robots to autonomous vehicles operating as mobility aids among pedestrians in unstructured environments. It demonstrates how agent interactions can be effectively modelled within prediction and planning modules, and how uncertainty in these predictions can be integrated into a DRL-based motion planner.