Situation Awareness for Autonomous Agents under Limited Sensing

Autonomous agents, such as robots and automated vehicles, rely on their ability to perceive and interpret their environment to make informed decisions and execute actions that align with their goals. A key aspect of this capability is situation awareness, which involves understanding the current state of the environment and predicting its future evolution. Traditional autonomous systems address perception and prediction as separate tasks within a sequential pipeline, where raw sensor data is processed into increasingly abstract representations. While this structured approach has driven significant advancements, it remains constrained by sensor limitations, including occlusions, measurement uncertainty, and adverse weather conditions. This thesis investigates how predictions from past observations can enhance perception algorithms, enabling agents to infer missing information, reduce uncertainty, and better anticipate changes. To support this integration, alternative environment representations are explored that allow feedback between prediction and perception while capturing uncertainty. This tighter coupling improves decision-making, particularly in complex and partially observable environments. The contributions include: (1) a reachability-based reasoning framework for tracking possible hidden obstacles; (2) its extension to handle delayed and partial external data; (3) a probabilistic mapping method, Transitional Grid Maps (TGM), that jointly models static and dynamic occupancy; and (4) an extension of TGM to mitigate weather-induced sensor noise. The proposed methods are evaluated in simulated and real scenarios where traditional perception pipelines struggle, such as occluded, highly dynamic and noisy environments. By bridging the gap between perception and prediction, this work contributes to the development of more robust and intelligent autonomous systems.