To Örebro University

Safe and efficient robot operation in human-shared environments requires understanding and anticipating how people move. This thesis advances the representation, learning, and exploitation of spatio-temporal Maps of Dynamics (MoDs), which are structured models that encode statistical motion patterns observed in an environment, for long-horizon human motion prediction and human-aware robot navigation.

On the representation side, this thesis addresses key limitations of existing MoDs along four complementary dimensions: adaptivity, continuity, robustness, and context-awareness. First, for adaptivity, an online learning method based on stochastic Expectation Maximization enables MoDs to adapt incrementally to changing motion patterns without costly offline reconstruction, maintaining accuracy at substantially lower computational cost and enablin glong-term deployment on mobile robots without storing historical data. Second, for continuity, a neural implicit representation eliminates spatial discretization by learning a continuous function that maps arbitrary spatio-temporal coordinates to multi-modal velocity distributions, achieving higher representation accuracy than grid-based approaches while training efficiently. Third, for robustness, a method inspired by data-driven airflow modeling extracts the structured laminar component from observed motion, mitigating the impact of anomalous data and producing more robust motion pattern representations. Fourth, for context-awareness, conditioning MoDs on time of day and agent class captures heterogeneous motion behaviours that single aggregated maps cannot represent, yielding more informative motion priors that directly improve downstream prediction accuracy.

On the exploitation side, MoDs are shown to serve as powerful environmentlevel priors for downstream tasks. This thesis exploits MoDs first f or longhorizon human motion prediction, where a sampling-based framework queries MoDs at each prediction step to generate multi-modal, spatially grounded trajectories that respect the environment topology over horizons of up to 60 seconds. The predicted trajectories implicitly follow corridors, turn at junctions, and reach semantically meaningful goal regions, improving interpretability. Across multiple datasets and prediction horizons, MoD-based methods outperform learning-based and diffusion-based baselines at long horizons, with the advantage being most pronounced under data scarcity and class imbalance. This framework is further extended with laminar enhancement, temporal conditioning, and agent-class conditioning, along with a trajectory ranking mechanism based on cumulative MoD likelihood that provides a principled way to select the most probable prediction from the multi-modal output, improving practical utility for real-world deployment.

This thesis also exploits MoDs for human-aware robot navigation. A hierarchical framework that treats crowd motion as guidance rather than solely as an obstacle enables proactive flow alignment in dense crowds, addressing the freezing robot problem and successfully transferring from simulation to real-world deployment in a public museum without retraining. Finally, MoD priors are integrated into multi-agent path finding to reduce conflicts between robots and human flows in shared environments, extending the utility of structured motion representations from single-robot planning to multi-robot coordination.

Together, these contributions demonstrate that learned structured representations of human dynamics provide a unifying foundation for prediction and navigation, extending the robot’s contextual awareness beyond its immediate sensing range toward more predictive, socially aware, and robust autonomous operation in human-shared spaces.