Autonomous Vehicle Trajectory Planning Using Virtual Doppelgangers

The system transitions from treating predicted trajectories as static hard obstacles to modeling them as dynamic, probabilistic modalities. The AV generates multiple possible trajectories (modalities) for other actors, each with associated probabilities, and updates these predictions as new information becomes available. This allows the AV to adapt its motion plan based on the evolving likelihood of different actor behaviors rather than avoiding all predicted paths as if they were certain.

The system changes the parameter representation of other actors' future positions from deterministic coordinates to probability distributions across multiple modalities. Each modality represents a distinct behavioral pattern (e.g., continuing straight, turning left, turning right) with an associated probability. This parameter transformation enables the AV to reason about uncertainty and make decisions based on the most likely scenarios while remaining prepared for alternatives.

The system segments the continuous space of possible actor reactions into a discrete set of modalities, each representing a distinct behavioral category. Instead of reasoning about all possible continuous trajectories, the AV identifies key modalities (e.g., maintain current behavior, accelerate, decelerate, change lane) and generates representative trajectories for each. This segmentation reduces the infinite complexity of continuous motion spaces to a manageable number of discrete categories that can be evaluated efficiently.

The system performs partial reasoning by focusing computational resources on the most probable modalities rather than exhaustively evaluating all possible actor reactions. The AV generates predictions for multiple modalities but prioritizes planning based on the highest-probability scenarios, using computational approximations and pruning techniques to avoid evaluating every possible interaction outcome. This partial action approach provides sufficiently good decisions without the full computational burden of complete enumeration.