Trajectory Ranker

Purpose/Role

This package provides a C++ library that scores candidate trajectories and selects the best one. Ranking combines three independently configurable terms:

1. Safety — penalty from the validator risk level (RiskLevel)
2. Source — penalty from the trajectory generator (diffusion vs backup planners)
3. Quality — optional metric-plugin evaluation (comfort / progress / consistency)

The library can be embedded into a calling ros node (such as autoware_trajectory_selector) via TrajectoryRankerWrapper.

Architecture

Algorithm Overview

For each call to TrajectoryRanker::process():

1. Safety evaluation (safety.enable): map each trajectory’s RiskLevel to a penalty via safety.levels / safety.penalty.
2. Source evaluation (source.enable): map each trajectory’s TrajectorySource to a penalty via source.levels / source.penalty.
3. Quality evaluation (evaluation.enable, default off): resample each trajectory relative to ego odometry, run the configured metric plugins, and set quality_penalty = 1 - quality_score. Skipped if route handler or odometry is unavailable.
4. Score aggregation: [ \mathrm{score} = 1 - \frac{ s{\mathrm{safety}}\,p{\mathrm{safety}}
   • s{\mathrm{source}}\,p{\mathrm{source}}
   • s{\mathrm{eval}}\,p{\mathrm{quality}} }{ s{\mathrm{safety}} + s{\mathrm{source}} + s{\mathrm{eval}} } ] where (s_) are the configured scales and (p__) are the per-trajectory penalties (zero when that term is disabled).
5. Best selection: the trajectory with the highest score is retained as best_trajectory_info and used to update trajectory history / previous points for subsequent quality metrics.

Trajectory sources

Source mapping from generator names is performed in the calling node (e.g trajectory_selector_node) before calling the ranker.

Default penalties (from config/trajectory_ranker.param.yaml)

Quality Metrics

Used only when evaluation.enable is true. Plugins are loaded by Evaluator from evaluation.plugin_names.

• Maximization: higher raw values are better
• Deviation: lower raw values are better

When quality evaluation is enabled, trajectories are resampled to evaluation.sampling_number points at evaluation.sampling_resolution seconds relative to the current ego pose.

Interface

This package is primarily a C++ library consumed by trajectory_selector_node. The selector owns the ROS I/O and passes:

• RankerInputTrajectories — each entry has a CandidateTrajectory, validator risk_level, and TrajectorySource
• RankerContext — odometry, route handler, and generator info (needed for quality metrics and output generator metadata)

Selector-side topics

Parameters

{{ json_to_markdown(“planning/autoware_trajectory_ranker/schema/trajectory_ranker.schema.json”) }}

Changelog for package autoware_trajectory_ranker

0.53.0 (2026-09-29)

• Merge remote-tracking branch 'origin/main' into prepare-0.53.0-changelog

• fix(planning): fix ENABLE_AGNOCAST=1 build and startup of trajectory_selector (#13364)

  • fix(trajectory_selector): fix ENABLE_AGNOCAST=1 build

  * build(trajectory_ranker): add autoware_agnocast_wrapper_setup to the library target ---------

• feat(trajectory_ranker): implement and integrate ranker into selector node (#13353)

  • feat(trajectory_ranker): implement new ranker module and integrate into selector component (#3208)
  • add trajectory_ranker_wrapper framework
  • refactor trajectory ranker parameter handling
  • implement trajectory_ranker class framework
  • implement core ranker logic
  • add logic to evaluate trajectories based on risk level
  • add logic to evaluate trajectories based on source
  • use existing metrics based evaluation to evaluate trajectory quality
  • integrate new ranker into trajectory_selector_node
  • refactor code
  • remove obsolete ranker node
  • refactor for debugging
  • add flag to enable/disable ranker within selectory node
  • fix parameter update logic, cleanup code
  • update launch files
  • disable quality evaluation by default
  • fix topic name
  • minor refactor
  • output debug to console when best trajectory has low score
  • support new backup planner dual go/stop trajectories
  • populate generator info of ScoredCandidateTrajectories
  • filter out shadow mode metrics before assigning combined trajectory risk level
  • update source penalties
  • add integration tests for trajectory ranker
  • add ranker parameters schema, update readme
  • remove simple_trajectory_ranker_node
  • remove launch prefix
  • pass active_filter_names to validator from wrapper

  * add missing includes ---------

  • replace rclcpp::Node usage by agnocast_wrapper::Node
  • fix selector node tests

  * add missing selector config file ---------

• refactor(planning): move node design files into each package (#13102) Co-authored-by: Claude Opus 5 (1M context) <<noreply@anthropic.com>>

• feat(simple_trajectory_ranker): apply agnocast_wrapper::Node to simple_trajectory_ranker (#13065)

  • feat: apply agnocast
  • fix; test

  * style(pre-commit): autofix ---------Co-authored-by: pre-commit-ci-lite[bot] <117423508+pre-commit-ci-lite[bot]@users.noreply.github.com>

• feat(diffusion_planner,trajectory_ranker,trajectory_adapter): populate turn_indicator field on CandidateTrajectory (#12922)

  • add turn_indicator into cantidate trajectory
  • rename field
  • fix rebase misstake
  • changes for turn_indicator topic publisher from adapter

  * fix turn_indicators_command timestamp ---------

• fix(diffusion_planner, trajectory ranker): remove builder pattern (#12944) remove builder pattern

• Contributors: Kotakku, Ryohsuke Mitsudome, Taekjin LEE, Yutaro Kobayashi, mkquda

0.52.0 (2026-06-30)

• Merge remote-tracking branch 'origin/main' into tmp/bot/bump_version_base

• feat(planning): apply autoware_agnocast_wrapper to diffussion

File truncated at 100 lines see the full file