Equi-RO: A 4D mmWave Radar Odometry via Equivariant Networks

Technical Summary of Equi-RO: A 4D mmWave Radar Odometry via Equivariant Networks

Problem Statement
Accurate odometry is critical for autonomous vehicles and robots, particularly in GPS-denied environments. While LiDAR and cameras are standard, they often fail under extreme weather conditions. 4D millimeter-wave (mmWave) radar offers a robust, all-weather alternative with velocity measurement capabilities. However, existing 4D radar odometry methods face two primary challenges:

1. Data Sparsity and Noise: 4D radar point clouds are significantly sparser and noisier than LiDAR data, making feature aggregation and correspondence matching difficult.
2. Doppler Velocity Coupling: Doppler velocity measurements depend on both the target's motion and the sensor's (ego-vehicle's) motion. This coupling means raw Doppler features are neither invariant nor equivariant under sensor transformations (specifically SE(3)), limiting the applicability of equivariant neural networks which rely on predictable feature transformations.

Methodology: Equi-RO
The authors propose Equi-RO, a framework that integrates equivariant neural networks with a specialized pre-processing pipeline to address the unique characteristics of 4D radar data. The methodology consists of three core components:

1. Invariant Doppler Pre-processing:

   • Ego-Velocity Estimation: An iterative reweighted least squares algorithm estimates the ego-velocity ($v_{ego}$) by minimizing Doppler residuals for stationary points.
   • Velocity Compensation: The raw Doppler velocity is compensated by subtracting the ego-motion component. This yields a compensated node velocity ($v'_{dop}$) that is invariant to SO(3) rotations of the sensor.
   • Edge Velocity: A compensated relative velocity is derived for edges between nodes. This feature is shown to be SE(3)-invariant when nodes belong to rigid objects with identical absolute velocities.

2. Graph Construction:

   • To mitigate sparsity, the point cloud is structured as a graph where each point is a node.
   • Edges are established based on a cost function combining spatial distance and velocity similarity, explicitly connecting kinematically consistent neighbors.
   • Features are categorized by their transformation properties:
     • Equivariant Node Features: 3D Euclidean coordinates.
     • Invariant Node Features: Compensated Doppler velocity, Radar Cross Section (RCS), and node degree.
     • Invariant Edge Features: Compensated relative velocity and Euclidean distance.

3. Equivariant Network Architecture:

   • Stage 1 (Geometric Extraction): An SO(3)-equivariant Dynamic Graph Convolutional Neural Network (DGCNN) using Lie Neurons (LN) processes the geometric coordinates. This stage focuses on learning rotation-consistent features after compensating for dominant translational motion.
   • Stage 2 (Feature Aggregation): An SE(3)-equivariant Graph Neural Network (EGNN) integrates the processed geometric features with the invariant physical features (velocity, RCS) via message passing. A Gaussian kernel is employed to transform equivariant features into invariant representations for message aggregation.
   • Transformation Estimation: The network outputs per-point feature sets for source and target frames. Correspondences are established via a learnable similarity metric and a differentiable Sinkhorn algorithm, followed by a weighted Singular Value Decomposition (SVD) to estimate the relative transformation $(R, t)$.
   • Loss Function: The model is trained end-to-end using a composite loss that includes rotation, translation, pitch, and yaw errors, with learnable balancing parameters to handle scale differences.

Key Contributions

1. Novel Algorithm: Development of a 4D radar odometry algorithm based on equivariant networks, specifically designed to maintain accuracy during large rotations (e.g., sharp turns).
2. Invariant Feature Derivation: A dedicated pre-processing pipeline that mathematically derives invariant Doppler features, decoupling object motion from ego-vehicle rotation to stabilize motion-related features.
3. Dual-Feature Framework: A network architecture that separately processes equivariant geometric structures and invariant physical attributes, preserving geometric consistency while leveraging radar-specific physical data.
4. Empirical Validation: Extensive experiments on the NTU4DRadLM dataset and a self-collected dataset, including ablation studies, demonstrating the method's effectiveness across diverse scenarios.

Experimental Results

• Performance on NTU4DRadLM: Equi-RO outperformed state-of-the-art baselines (including ICP variants, APDGICP, Radar4Motion, RaFlow, and CAO-RONet). Compared to the best baseline, it achieved 10.7% improvement in translation accuracy and 13.4% improvement in rotation accuracy on average.
• Large Rotation Scenarios: The method demonstrated superior robustness in segments with sharp turns and large rotations, where traditional methods often degraded significantly.
• Generalization: When tested on a self-collected dataset with lower-resolution radar (trained only on NTU4DRadLM), Equi-RO maintained the best translation and rotation accuracy, indicating strong cross-dataset generalization.
• Ablation Studies: Removing the invariant velocity compensation or the equivariant network design resulted in significant performance degradation, confirming the necessity of both components. The variant without equivariant design showed the worst results, highlighting the critical role of the equivariant architecture.

Significance and Claims
The paper claims that Equi-RO provides a new perspective for processing 4D radar point clouds by effectively combining equivariant learning with physical attribute integration. The authors assert that their method successfully addresses the limitations of existing approaches in handling sparse, noisy data and large rotational variations. While the method achieves real-time performance (58.15 ms/frame on an NVIDIA A100), the authors modestly note that further optimization (e.g., C++ migration, TensorRT) is required for deployment on embedded on-board platforms. The work highlights the potential of equivariant networks for radar-based perception and suggests future directions involving multi-modal fusion with IMUs and cameras.