跳转至

Point-LIO Handoff Context

Date: 2026-04-16 Workspace: ~/venom_ws/src/venom_vnv Primary package: ~/venom_ws/src/venom_vnv/localization/lio/Point-LIO Primary bringup: ~/venom_ws/src/venom_vnv/venom_bringup

User Goal

The user wants to investigate and improve Point-LIO performance, especially drift during high-speed rotation and high-speed motion, and then run controlled tests using recorded raw Mid360 bags.

Repository / Branch Context

  • Current Point-LIO remote setup:
  • origin: git@github.com:Venom-Algorithm/Point-LIO.git
  • upstream: https://github.com/hku-mars/Point-LIO.git
  • Current checked-out commit in localization/lio/Point-LIO during this work:
  • 568f60c6179cdb1455ea3739bf7de76147466fbc
  • HKU upstream comparison target used:
  • upstream/point-lio-with-grid-map

What Was Read / Analyzed

  • Top-level project architecture and docs were read.
  • Entire localization/lio/Point-LIO tree was read, including:
  • src/*
  • config/*
  • launch/*
  • include/*
  • bundled third-party headers (IKFoM, ivox, matplotlibcpp)
  • The actual user test launch was identified as:
  • ~/venom_ws/src/venom_vnv/venom_bringup/launch/examples/mid360_point_lio.launch.py
  • That launch loads this YAML, not the package-internal default:
  • ~/venom_ws/src/venom_vnv/venom_bringup/config/examples/point_lio_mapping.yaml

Important Technical Findings

1. Fixed time offset was not actually wired through

The bringup YAML already had:

  • ~/venom_ws/src/venom_vnv/venom_bringup/config/examples/point_lio_mapping.yaml
  • common.time_diff_lidar_to_imu: 0.0

The package also read this parameter in:

  • ~/venom_ws/src/venom_vnv/localization/lio/Point-LIO/src/parameters.cpp

But before this session it was not effectively connected into the IMU timestamp correction path for the current branch.

HKU upstream does not contain a robust automatic millisecond-grade LiDAR-IMU time offset estimator for this use case. Instead, upstream supports a fixed configured offset and applies it in the IMU callback.

2. Mid360 synchronization conclusion from user testing

The user tested fixed time offset tuning and concluded:

  • time_diff_lidar_to_imu = 0.0 works best
  • Mid360 appears to already provide good internal synchronization in this setup

Therefore:

  • Do not prioritize adding an automatic LiDAR-IMU time offset estimator next
  • Keep fixed offset support working, but leave the effective value at 0.0

3. Real high-priority bug found in current branch

In laserMapping.cpp, vectors were using reserve() and then being written with operator[], which is unsafe and can cause out-of-bounds writes:

  • crossmat_list
  • pbody_list

This is especially suspicious under high dynamic motion because it can manifest as unstable behavior or unexplained drift.

Code Changes Already Made

A. Fixed vector sizing bug

File: - ~/venom_ws/src/venom_vnv/localization/lio/Point-LIO/src/laserMapping.cpp

Change: - Replaced: - crossmat_list.reserve(feats_down_size); - pbody_list.reserve(feats_down_size); - With: - crossmat_list.resize(feats_down_size); - pbody_list.resize(feats_down_size);

B. Wired fixed LiDAR->IMU time offset into IMU callback

File: - ~/venom_ws/src/venom_vnv/localization/lio/Point-LIO/src/li_initialization.cpp

Change: - IMU timestamp correction now subtracts time_diff_lidar_to_imu in imu_cbk()

Current relevant line area: - ~/venom_ws/src/venom_vnv/localization/lio/Point-LIO/src/li_initialization.cpp:160

C. Added startup log for configured fixed time offset

File: - ~/venom_ws/src/venom_vnv/localization/lio/Point-LIO/src/parameters.cpp

Change: - Added startup log showing the active fixed LiDAR->IMU offset

Current relevant line area: - ~/venom_ws/src/venom_vnv/localization/lio/Point-LIO/src/parameters.cpp:303

Bringup / Recording Changes Already Made

Added raw Mid360 recording launch

File: - ~/venom_ws/src/venom_vnv/venom_bringup/launch/examples/mid360_record_raw.launch.py

Purpose: - Start livox_ros_driver2 - Record only raw sensor + TF topics - Save bags in a persistent user directory

Default output directory: - ~/venom_bags/mid360_raw/

Recorded topics: - /livox/lidar - /livox/imu - /tf - /tf_static

Deleted unneeded launch

Deleted file: - ~/venom_ws/src/venom_vnv/venom_bringup/launch/examples/mid360_point_lio_record.launch.py

This was removed because the user explicitly said it was not needed.

Commands Given To User

Record raw Mid360 bag

source ~/venom_ws/src/venom_vnv/install/setup.bash
ros2 launch venom_bringup mid360_record_raw.launch.py

Record raw Mid360 bag to a named directory

source ~/venom_ws/src/venom_vnv/install/setup.bash
ros2 launch venom_bringup mid360_record_raw.launch.py bag_dir:=/Users/liyh/venom_bags/mid360_raw/high_speed_turn_01

Replay a raw bag

source ~/venom_ws/src/venom_vnv/install/setup.bash
ros2 bag play /Users/liyh/venom_bags/mid360_raw/high_speed_turn_01 --clock

Start current Point-LIO bringup for testing

source ~/venom_ws/src/venom_vnv/install/setup.bash
ros2 launch venom_bringup mid360_point_lio.launch.py

User's Current Testing Strategy

The user agreed that the right next step is to record several raw bags covering motion intensity levels, then replay the same raw bags repeatedly while changing Point-LIO parameters.

Suggested bag set:

  1. low_speed_01
  2. mid_speed_01
  3. high_speed_turn_01
  4. high_speed_translate_turn_01

Recording guidance already discussed:

  • Each bag should be separate
  • Include a few seconds of stillness at start and end
  • Prefer 20 to 60 seconds each
  • Goal is fair A/B comparison from the same raw input

Key Performance / Drift Hypotheses To Investigate Next

These were the most relevant likely causes after time offset was ruled out as the main issue:

  1. Real-time load under aggressive settings
  2. User's active bringup YAML is more aggressive than package default
  3. Especially:
    • ~/venom_ws/src/venom_vnv/venom_bringup/config/examples/point_lio_mapping.yaml:9
    • filter_size_surf: 0.05

  4. This may improve geometry but can hurt real-time behavior at high speed

  5. High-dynamic IMU / propagation quality

  6. Current IMU_Processing.cpp is relatively light and mostly initialization + passthrough

  7. More explicit or refined high-dynamic handling may be needed later

  8. Parameter set mismatch for high dynamic motion

  9. Need proper replay-based A/B testing, not live tuning

  10. Extrinsic sensitivity under vibration / high angular motion

  11. extrinsic_est_en is currently False, which is consistent with upstream recommendations for aggressive motion if extrinsics are known
  1. Do not spend more time on automatic LiDAR-IMU fixed time offset estimation unless new evidence appears.
  2. Use the new raw-record launch to build a reproducible bag suite.
  3. Create a structured replay-based test plan using the user's actual bringup:
  4. ~/venom_ws/src/venom_vnv/venom_bringup/launch/examples/mid360_point_lio.launch.py
  5. Compare parameter sets on the same raw bags.
  6. Prioritize profiling / timing visibility next.

Recommended immediate work items:

  1. Add lightweight profiling output for:
  2. preprocess time
  3. propagation time
  4. update time
  5. map incremental time
  6. total frame time
  7. Create a dedicated high-dynamic YAML variant alongside:
  8. ~/venom_ws/src/venom_vnv/venom_bringup/config/examples/point_lio_mapping.yaml
  9. Evaluate whether filter_size_surf: 0.05 is too aggressive for stable real-time performance during high-speed rotation.

Files Most Relevant For Follow-Up

  • ~/venom_ws/src/venom_vnv/venom_bringup/launch/examples/mid360_point_lio.launch.py
  • ~/venom_ws/src/venom_vnv/venom_bringup/launch/examples/mid360_record_raw.launch.py
  • ~/venom_ws/src/venom_vnv/venom_bringup/config/examples/point_lio_mapping.yaml
  • ~/venom_ws/src/venom_vnv/localization/lio/Point-LIO/src/laserMapping.cpp
  • ~/venom_ws/src/venom_vnv/localization/lio/Point-LIO/src/li_initialization.cpp
  • ~/venom_ws/src/venom_vnv/localization/lio/Point-LIO/src/parameters.cpp
  • ~/venom_ws/src/venom_vnv/localization/lio/Point-LIO/src/IMU_Processing.cpp