Towards Explaining Uncertainty Estimates in Point Cloud Registration

TL;DR

This work addresses the lack of interpretability in ICP-derived uncertainty by introducing kernel SHAP to attribute pose uncertainty to three tangible sources: sensor noise, initial pose uncertainty, and partial overlap. It combines perturbation-based data generation with KL-divergence to quantify uncertainty and uses SHAP values to explain how each source contributes to the observed uncertainty in ICP outputs. The experiments show that sensor noise often dominates, with partial overlap also playing a significant role, and demonstrate interpretable explanations through summary, waterfall, and dependence plots. The approach is model-agnostic and has practical implications for active perception and teleoperation, enabling systems to understand and mitigate failure modes in real time.

Abstract

Iterative Closest Point (ICP) is a commonly used algorithm to estimate transformation between two point clouds. The key idea of this work is to leverage recent advances in explainable AI for probabilistic ICP methods that provide uncertainty estimates. Concretely, we propose a method that can explain why a probabilistic ICP method produced a particular output. Our method is based on kernel SHAP (SHapley Additive exPlanations). With this, we assign an importance value to common sources of uncertainty in ICP such as sensor noise, occlusion, and ambiguous environments. The results of the experiment show that this explanation method can reasonably explain the uncertainty sources, providing a step towards robots that know when and why they failed in a human interpretable manner

Figures (4)

• Figure 1: The simplified input on the left, represented by binary conditions, are mapped to the feature values on the right via the mapping function $h_x$.
• Figure 2: SHAP summary plot.
• Figure 3: SHAP waterfall plot.
• Figure 4: SHAP dependence plots.