Autonomous Vehicle Trajectory Planning for Route-Aware Mode Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous driving methods, such as the End-to-End Interpretable Neural Motion Planner (NMP), generate trajectories without considering the desired route and marginalize possible ego intentions and scene developments, leading to inefficient and potentially unsafe driving behaviors.
Innovation Solution
A method that involves obtaining sensor data, evaluating potential vehicle states using a first evaluation function, identifying local optima, generating candidate trajectories, evaluating state transitions with a second evaluation function, and selecting optimal trajectories that consider desired routes and mode differentiation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If NMP generates trajectory solely on basis of sensor data and map information, then collision-free driving is achieved, but desired route is not considered
Solution Approach 1:
The patent combines the NMP cost volume approach with PGP graph neural network predictions to create a unified trajectory planning system. The cost volume provides collision-free path information while the PGP graph incorporates desired route and navigation information, merging both approaches to achieve both safety and route adherence.
Solution Approach 2:
The patent introduces a trajectory evaluation module as an intermediary that assesses candidate trajectories against multiple criteria including collision freedom, route alignment, and comfort. This mediator selects the optimal trajectory that satisfies both safety requirements and desired route constraints.
2Productivity
If NMP generates single cost volume or single prediction, then computational efficiency is maintained, but marginalization of possible ego intentions and scene developments occurs
Solution Approach 1:
The patent segments the trajectory planning into multiple independent prediction branches within the PGP framework, each representing different possible ego intentions and scene developments. These segmented predictions are then aggregated to form a comprehensive view without requiring generation of all possible trajectories explicitly.
Solution Approach 2:
The patent generates a limited set of diverse candidate trajectories that cover the most important possible intentions and scene developments, rather than exhaustively generating all possible trajectories. This partial action approach maintains computational efficiency while capturing essential uncertainty.
3Ease of manufacture
If NMP selects trajectory from randomly generated candidates, then implementation is simple, but trajectory quality depends on candidate set and may be significantly worse than optimum
Solution Approach 1:
The patent performs preliminary generation of diverse candidate trajectories using PGP graph neural networks before the final selection stage. This preliminary action creates a richer set of informed candidates that are more likely to include high-quality trajectories, improving the quality of the final selected trajectory while maintaining the simple selection framework.
4Adaptability or versatility
If NMP generates highly complex trajectories with S-curves, then trajectory coverage is improved, but mode mixing occurs and impermissible driving behavior is selected
Solution Approach 1:
The patent implements a trajectory evaluation and selection module that provides feedback on the quality and permissibility of candidate trajectories. This feedback mechanism identifies and rejects trajectories with impermissible mode mixing while selecting those that appropriately use complex maneuvers like S-curves when genuinely needed for safe navigation.
Data Source
AI summary
A method for planning an optimal driving behavior for an at least partially autonomously driving vehicle. The method includes: obtaining sensor data of the vehicle; ascertaining a first evaluation function which assigns a quality to each possible state of the vehicle at discrete points in time within a planning horizon of the vehicle based on the sensor data; ascertaining a local optimum of the first evaluation function at each discrete point in time; ascertaining a candidate trajectory which includes a temporal sequence of the local optima; evaluating the candidate trajectory using a second evaluation function which evaluates state transitions between successive states of the at least one candidate trajectory; selecting an optimal candidate trajectory from the candidate trajectories based on the first and second evaluation functions; and transmitting the selected at least one optimal candidate trajectory to a control unit of the vehicle.