Autonomous Vehicle Trajectory Planning for Reactive Road User Forecasting

The method segments the exponentially large space of possible trajectories into a manageable set of representative trajectories. Instead of reasoning over all possible combinations of actor actions, the system identifies and focuses on a segmented subset of key trajectories that capture the essential interaction patterns, thereby reducing computational complexity while maintaining prediction accuracy.

The system introduces an intermediary representation in the form of a probabilistic interaction model that mediates between the AV's actions and other actors' reactions. This intermediary model captures the essential dynamics of actor interactions without requiring exhaustive enumeration of all possible trajectories, enabling efficient reasoning about reactions.

The system transforms the static concept of hard obstacles into dynamic probabilistic constraints. Instead of treating other actors' future positions as fixed boundaries, the method models them as probabilistic distributions that evolve based on the AV's planned actions. This dynamic approach maintains planning tractability while enabling adaptive decision-making that considers how other actors may react to the AV's movements.

The method changes the parameter representation of other actors from deterministic position coordinates to probabilistic trajectory distributions. By representing future positions as probability distributions rather than fixed points, the system maintains computational simplicity while capturing the uncertainty and reactiveness of other actors, leading to superior decision-making.