Abstract
Classical robotics algorithms like Iterative Closest Point (ICP) are interpretable and modular but often struggle with noise and poor initialization. Deep learning approaches offer robustness but typically operate as "black boxes." This paper introduces NAR-*ICP, a framework that applies Neural Algorithmic Reasoning (NAR) to point cloud registration. By training Graph Neural Networks (GNNs) to execute the step-by-step logic of ICP algorithms using intermediate "hints," NAR-ICP achieves superior accuracy and robustness compared to traditional methods. The resulting models are lightweight, interpretable, and capable of surpassing the performance of the classical algorithms they were trained to mimic.
Introduction
In robotics, point cloud registration—the process of aligning two 3D datasets—is a fundamental task. The trade-off between classical algorithms and neural networks is well-documented: classical ICP is predictable but fragile, while neural networks are robust but opaque. NAR-ICP addresses this by creating a hybrid system where a neural network learns to "reason" through the intermediate computational steps of a proven robotics algorithm.
Methodology
The NAR-ICP framework extends the CLRS-30 benchmark into the robotics domain, representing point clouds as dynamic bipartite graphs where nodes are points and edges represent potential correspondences.
Architecture
The system follows an Encoder-Processor-Decoder structure:
- Encoders: Transform raw 3D coordinates and semantic labels into latent embeddings.
- Processor: A recurrent GNN (specifically a Triplet-MPNN) that performs iterative reasoning steps.
- Decoders: Translate latent states back into predicted point positions and correspondences.
Algorithmic Hints and Termination
A core component of the NAR approach is the use of algorithmic hints. During training, the network is supervised on the intermediate states of a classical ICP algorithm, such as current correspondences, distances, and error metrics. This guides the latent space to follow a logical trajectory.
Furthermore, the model includes a learned termination network. Unlike standard NAR models that run for a fixed number of iterations, this network learns to detect convergence, significantly improving inference speed by reducing the number of required steps by over 2x.
Experimental Evaluation
The framework was evaluated on both synthetic and real-world datasets (SemanticKITTI).
Key Results
- Accuracy: NAR-ICP variants significantly outperformed classical baselines. On SemanticKITTI, the NAR-P2L+ model achieved a Relative Translation Error (RTE) of 0.135m, a massive improvement over the 1.032m error of classical P2L-ICP.
- Efficiency: The model is highly optimized, featuring only 773k parameters—orders of magnitude smaller than many state-of-the-art deep learning registration models like DGR (235M parameters).
- Robustness: By learning to infer an optimal initial alignment, NAR-ICP avoids the local minima that often plague classical ICP when starting from large offsets.
Discussion and Conclusion
NAR-ICP demonstrates that neural networks can "surpass their teacher." By training on both successful algorithmic steps and adversarial examples where classical ICP fails, the neural execution becomes more resilient. The use of hints makes the model's "trajectories" interpretable, allowing researchers to monitor convergence.
Future work will focus on scaling this approach to dense geometric features and integrating NAR-ICP into full-scale SLAM and path-planning systems for real-time robotics applications.
Summary updated