[vividf]

Description

Make the point cloud concatenation core (CombineCloudHandler, CombineCloudHandlerBase,
the concatenation diagnostics) independent of the ROS runtime, so it can be driven from a
non-node context (the offline/Python pipeline in #13113 / #12874) and unit-tested without
spinning a node. The node keeps exactly the same topics, parameters and outputs.

This PR is the merged result of a 4-part stack that was reviewed step by step on the fork
(links under Related links); each step is still a separate commit here.

1. Type the matching strategy — MatchingStrategyType enum pinned to the
   ConcatenatedPointCloudInfo STRATEGY_* constants; the string is parsed once, at the
   parameter boundary (parse_matching_strategy()); matching_strategy_name_map is gone.
2. Key the handler on the message type — CombineCloudHandler<sensor_msgs::msg::PointCloud2>
   / CombineCloudHandler<cuda_blackboard::CudaPointCloud2> instead of the *Traits tag types.
3. Drop pcl and rclcpp time from the core — stamps are builtin_interfaces::msg::Time
   with a small time_utils.hpp (is_earlier, add, subtract, to_seconds);
   pcl_ros::transformPointCloud / pcl::concatenatePointCloud are replaced by an Eigen
   transform loop and a byte append on the shared XYZIRC layout.
4. Inject transforms instead of holding a node — the handler no longer owns a
   ManagedTransformBuffer; the node looks up sensor→output extrinsics and calls
   set_transform(). A source whose frame has no transform is dropped and reported
   (dropped_sources_missing_transform, STATUS_INVALID) instead of being placed with an
   identity transform. Motion-compensation warnings became a MotionCompensationStatus the
   node logs.
5. Build the diagnostics outside the node — build_diagnostic_status() in
   concatenation_diagnostics.{hpp,cpp} produces the DiagnosticStatus from a plain summary
   struct; the node only feeds it and publishes. A source without a transform now counts as
   not concatenated in the diagnostics, consistent with ConcatenatedPointCloudInfo.

A new concatenate_core library target contains only the ROS-runtime-free sources and links
message targets only, to keep it that way.

Related links

• Previous PR (merged): refactor(autoware_pointcloud_preprocessor): extract matching policy and rename collector matcher #12872 — extract matching policy and rename collector matcher
• Reviewed as a stack on the fork, now merged into this branch:
  refactor(autoware_pointcloud_preprocessor): key the combine handler on the point cloud message type vividf/autoware.universe#4, feat: add simple planning simulator package #5, feat:add common math pkg #6, feat: add autoware_utils package #7
• Follow-ups that build on this core: feat(autoware_pointcloud_preprocessor): add a ROS-runtime-free concatenator that owns the whole collector lifecycle #13113 (feat/concat_core), feat(autoware_pointcloud_preprocessor): add Python bindings for the concatenation core #12874 (Python bindings)

How was this PR tested?

• colcon build/test --packages-select autoware_pointcloud_preprocessor autoware_cuda_pointcloud_preprocessor
  — 519 tests, 0 failures (incl. lint).
  • test_concatenation_info: typed strategy.
  • test_concatenation_diagnostics (new, 11 cases): key order, level/message rules, missing
    topic, missing transform, late drop, advanced window.
  • test_cuda_combine_cloud_handler (new, 2 cases, runs on the GPU and self-skips without one):
    concatenation on device, drop of a source without an extrinsic.
  • test_concatenate_node_unit: transforms injected via set_transform() instead of a TF broadcaster.
• On-hardware A/B against main (same base commit, x86 + RTX, 3 lidars, motion compensation on,
  ~120 frames): output clouds equivalent (max |Δxyz| 23 µm, float32 rounding; identity-rotation
  sensor byte-identical); ConcatenatedPointCloudInfo identical; /diagnostics level identical on
  every frame; processing time −8 % median.

Notes for reviewers

• Review commit by commit if the whole diff is too much; the five steps above map to the commits.
• Unknown matching_strategy.type now throws std::invalid_argument from parse_matching_strategy()
  instead of the node's std::runtime_error. Both abort startup.
• transform_pointcloud() applies Matrix4f * Vector4f where pcl used linear()*v + translation();
  results differ only by float32 rounding (≤ 1–8 ULP, zero-mean).

Interface changes

None. No topic, parameter or message changes. Internal C++ API of the handler changed
(constructor no longer takes a rclcpp::Node; set_transform() added).

Effects on system behavior

• Nominal operation: none observable (verified as above).
• Missing TF for a source: previously the source was concatenated with an identity transform
  and source_info.status stayed OK. Now its points are dropped, source_info.status = INVALID,
  concatenation_success = false, and /diagnostics reports ERROR with
  Concatenated: <topic> = False plus a new key Frames without a transform to the output frame.
• One diagnostics message string changed: "dropped due to missing topics and its timestamp is
  earlier" → "... and because its timestamp is earlier". Exact-match monitoring rules need
  updating.
• Diagnostics timestamps are rendered from sec + nanosec*1e-9 (single rounding) instead of
  rclcpp::Time::seconds(); values may differ in the last digits.
• Misconfigured matching_strategy.type: different exception type/message, still aborts startup.

[github-actions]

Thank you for contributing to the Autoware project!

🚧 If your pull request is in progress, switch it to draft mode.

Please ensure:

• You've checked our contribution guidelines.
• Your PR follows our pull request guidelines.
• All required CI checks pass before marking the PR ready for review.

[sasakisasaki]

Quick Comment

I agree with merging this PR as the reasons below 👍

Reasons for Approval

• Characterization test passed: this indicates no regression for both core logic and ROS node logic
• I could tested ROS node free concatenation by using this example

I believe this PR certainly separates core logic from ROS node 🎉

CC: @vividf @mojomex

[drwnz]

Nice work. Some changes suggested. I did not test CUDA or Agnocast as both aren't working in my environment right now.

[sasakisasaki]

@drwnz Thank you for your review! That are make sense. Let me apply the fix based on your comments, respectively 👍

[sasakisasaki]

Now let me do double check if all the review comment is discussed tomorrow again.

[sasakisasaki]

@drwnz I applied fixes 👍

Let's wait review by @mojomex and @vividf 👍

[vividf]

@drwnz @mojomex @sasakisasaki
Thanks for the reviews, and @sasakisasaki for implementing the fixes!
@drwnz, could you take another look when you have time :)

[drwnz]

Thanks for the changes, LGTM!