Skip to main content

Testing MAD-ICP Robustness in CARLA: Odometry in Dynamic, Unfiltered Scenes

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

After my previous experiments with camera-based detection and FLOAM-based LiDAR odometry in CARLA, I wanted to push things further — this time focusing on robustness under adversarial conditions.

What is MAD-ICP?​

MAD-ICP (It's All About Matching Data) is a LiDAR odometry method built around three core ideas:

  • Minimal assumptions about the environment
  • Strong robustness across different motion profiles
  • Clean ICP-based formulation with smart data representation

What stood out to me is its philosophy: instead of relying on handcrafted features or heuristics, it focuses on better data association and uncertainty-aware mapping. The PCA-based point structure and keyframe selection based on pose quality are central to why it handles noise well.

Restructuring for ROS 2​

As with FLOAM, I simplified the original implementation to fit my pipeline:

  • Extracted the core algorithm into a pure C++ library (mad_icp_core)
  • Removed unnecessary components like Python bindings
  • Wrapped it with a ROS 2 node and connected it to CARLA

This keeps the system modular, clean, and easy to integrate or swap out for future experiments.

Deliberately Making the Problem Harder​

ICP-based methods are notoriously sensitive to dynamic objects. Rather than making things easy, I set up a stress test:

  • Spawned multiple moving vehicles in CARLA
  • Kept all LiDAR points from dynamic objects — no filtering
  • No object detection
  • No semantic understanding

The odometry had to deal with raw, contaminated point clouds by design. This was specifically intended to probe whether MAD-ICP is as robust as the paper claims.

Results​

The results were surprisingly good:

  • ✅ No visible map skew or distortion
  • ✅ No obvious gaps or misalignment
  • ✅ Without loop closure, the trajectory stayed consistent and returned close to the starting point

This is particularly striking because traditional ICP pipelines tend to degrade quickly in dynamic scenes — yet MAD-ICP maintained stable localization without any explicit dynamic object handling.

What This Experiment Reinforced​

Three observations stood out:

1. Data representation matters more than complexity. MAD-ICP's PCA-based structure and matching strategy handle noise well without requiring additional preprocessing or segmentation modules.

2. Uncertainty-aware mapping is powerful. Keyframe selection based on pose quality likely plays a significant role in filtering out the influence of dynamic points on the map.

3. Robustness doesn't always require more modules. Even without loop closure, object filtering, or semantic understanding, the system performed reliably. Sometimes doing the fundamentals — data association and uncertainty — exceptionally well is enough.

A well-designed classical method can still be extremely competitive. This experiment is a good reminder that algorithmic elegance and principled design often outperform complexity for its own sake.