Skip to main content

LiDAR Odometry in CARLA: Restructuring FLOAM for ROS 2

· 2 min read
Yi-Chen Zhang
Yi-Chen Zhang
Lead Engineer, AI and Autonomous

After my previous CARLA experiment with camera-based object detection, I wanted to push further — this time focusing on LiDAR and localization.

In the video you can observe:

  • The system performing LiDAR odometry in CARLA
  • Map misalignment during sharp turns — expected without loop closure, and likely addressable by integrating a loop closure module

Restructuring FLOAM for ROS 2​

The original FLOAM implementation does not support ROS 2. I restructured it into a cleaner, more modular architecture:

  • floam_core: a pure C++ library containing the core algorithm — no ROS dependency
  • floam: a ROS 2 node wrapping around floam_core

This separation makes the system modular and reusable — the core algorithm can be tested and developed independently of the middleware layer. On top of that:

  • Message synchronization is handled using ROS 2 message_filters
  • Subscribers and publishers are implemented in a multi-threaded design to maximize throughput

The performance improvement was noticeable:

ImplementationOdometry Update Time
Original FLOAM~70–80 ms
My ROS 2 implementation~30–40 ms

Simulating a Realistic LiDAR​

To make the simulation more realistic, I reconfigured CARLA's LiDAR model based on the Velodyne HDL-64E user manual. The resulting point clouds look surprisingly close to real sensor data — though visible discrepancies between simulated and real-world LiDAR remain.

I'm currently unsure about the best approach for a proper sim-to-real analysis for LiDAR. If anyone has experience or suggestions, I'd really appreciate the insights — feel free to reach out or leave a comment.

Takeaway​

This is another small but meaningful experiment showing that CARLA can be a practical tool for validating perception and localization pipelines before moving to real-world data. Between this and the camera-based detection experiment, CARLA is proving to be a genuinely useful sandbox for early-stage development.

The sim-to-real gap for LiDAR remains an open question for me — curious to hear how others are approaching it.