autoware_trajectory_ranker

Purpose

The autoware_trajectory_ranker package provides a flexible and extensible framework for evaluating and ranking multiple trajectory candidates in autonomous driving scenarios. It transforms a set of viable candidate paths into an ordered list by attaching a scalar score to every trajectory based on safety, comfort, and efficiency criteria.

This decoupling of upstream path-generation logic from downstream motion-control logic enables:

• Diverse trajectory generators (rule-based, ML-based, optimization-based) to coexist
• Controllers to simply follow the highest-ranked path
• Improved overall driving performance through multi-criteria optimization
• Architectural flexibility through a plugin-based metric system

Features

• Plugin-based Architecture: Dynamically load and configure evaluation metrics via pluginlib
• Multi-criteria Evaluation: Score trajectories based on multiple metrics simultaneously
• Time-based Weighting: Apply temporal decay to metrics for better near-term decision making
• Extensible Framework: Easy to add new metrics through the plugin system
• Real-time Processing: Efficient evaluation of multiple trajectories per planning cycle

Algorithm Overview

At every planning period the node latches the most recent odometry and perception context, then processes trajectories through five distinct stages:

1. Ingestion: Receives candidate trajectories from upstream generators
2. Resampling: Converts each path into a common frame and resamples to fixed resolution
3. Evaluation: Applies configured metric plugins to compute cost vectors
4. Aggregation: Combines metrics using weights and temporal decay
5. Publication: Outputs scored trajectories and debug information

Trajectory Resampling

Every incoming trajectory is re-interpolated relative to the ego pose so that:

• First point aligns with current vehicle position
• Path comprises exactly sample_num points
• Points are spaced at fixed temporal intervals
• All metrics operate on consistent trajectory representations

This ensures that metric plugins can assume identical temporal spacing even when generators produce paths of different lengths or densities.

Metric Evaluation

For each resampled path, the evaluator executes a plugin chain specified in the configuration. Each metric plugin:

• Computes a vector of values (one per trajectory point)
• Returns costs reflecting safety, comfort, or efficiency considerations
• Can be classified as either maximization or deviation metrics

Score Aggregation

The aggregation stage:

1. Applies per-metric weights from configuration
2. Multiplies by temporal decay weights (exponential decay for near-term bias)
3. Normalizes across trajectory candidates
4. Sums to produce a single scalar score per trajectory

Available Metrics

The package includes the following built-in metric plugins:

Metric Types

• Maximization: Higher values are better (e.g., distance traveled, time to collision)
• Deviation: Lower values are better (e.g., lateral acceleration, jerk)

Input / Output

Input Topics

Output Topics

Parameters

Core Parameters

File truncated at 100 lines see the full file