Human-Interpretable Uncertainty Explanations for Point Cloud Registration
Johannes Albert Gaus, Loris Schneider, Yitian Shi, Jongseok Lee, Rania Rayyes, Rudolph Triebel
AI summary
Problem
Standard registration methods like ICP fail under sensor noise, poor initialization, or occlusion, but existing uncertainty quantification techniques lack human-interpretable explanations of failure causes, hindering effective recovery.
Approach
GP-CA embeds aligned point clouds with a DGCNN and uses a multi-class Gaussian Process classifier to attribute uncertainty to semantic concepts, augmented by an active learning loop for online adaptation to new failure modes.
Key results
- Outperforms state-of-the-art uncertainty quantification and XAI methods in attribution accuracy and runtime efficiency
- Achieves high sample efficiency via an active learning loop that adapts to new uncertainty concepts with minimal labeling
- Demonstrates successful real-world robotic failure recovery by identifying occlusion and triggering targeted viewpoint changes
- Validates DGCNN embeddings and a score-then-diversify active learning strategy through extensive ablation studies
Why it matters
Provides actionable, interpretable failure explanations that enhance the robustness and autonomy of robotic perception systems in dynamic environments.
Abstract
In this paper, we address the point cloud reg- istration problem, where well-known methods like ICP fail under uncertainty arising from sensor noise, pose-estimation errors, and partial overlap due to occlusion. We extend prior work and develop a novel approach, Gaussian Process Concept Attribution (GP-CA), which not only quantifies registration uncertainty but also explains it by attributing uncertainty to well-known sources of errors in registration problems. Our approach leverages active learning to discover new uncertainty sources in the wild by querying informative instances. We validate GP-CA on three publicly available datasets and in our real-world robot experiment. Extensive ablations substantiate our design choices. Our approach outperforms other state-of- the-art methods in terms of runtime, high sample-efficiency with active learning, and high accuracy. Our real-world ex- periment clearly demonstrates its applicability under realistic multi-factor failure scenarios. Our video also demonstrates that GP-CA enables effective failure-recovery behaviors, yielding more robust robotic perception.