Angle of Arrival Estimation with Transformer: A Sparse and Gridless Method with Zero-Shot Capability

TL;DR

AAETR introduces a DETR-inspired transformer for gridless AOA estimation in automotive MIMO radars, achieving high-resolution angle estimates with sparse outputs. Trained on a large physics-based synthetic radar dataset, it outputs up to $M$ targets as $(\tilde{\boldsymbol{\uptheta}}, \tilde{\boldsymbol{\upalpha}}, \tilde{\mathbf{p}})$ via a bipartite matching loss, enabling end-to-end perception integration. Across extensive simulations and zero-shot real-data tests, AAETR outperforms traditional grid-based methods like IAA, delivering stronger sidelobe suppression, richer BEV representations, and robust performance under varying SNRs and target counts. The work demonstrates scalable gridless AOA with practical impact for ADAS/AV perception, offering a drop-in module capable of sim-to-real transfer and efficient, learnable angle estimation.

Abstract

Automotive Multiple-Input Multiple-Output (MIMO) radars have gained significant traction in Advanced Driver Assistance Systems (ADAS) and Autonomous Vehicles (AV) due to their cost-effectiveness, resilience to challenging operating conditions, and extended detection range. To fully leverage the advantages of MIMO radars, it is crucial to develop an Angle of Arrival (AOA) algorithm that delivers high performance with reasonable computational workload. This work introduces AAETR (Angle of Arrival Estimation with TRansformer) for high performance gridless AOA estimation. Comprehensive evaluations across various signal-to-noise ratios (SNRs) and multi-target scenarios demonstrate AAETR's superior performance compared to super resolution AOA algorithms such as Iterative Adaptive Approach (IAA). The proposed architecture features efficient, scalable, sparse and gridless angle-finding capability, overcoming the issues of high computational cost and straddling loss in SNR associated with grid-based IAA. AAETR requires fewer tunable hyper-parameters and is end-to-end trainable in a deep learning radar perception pipeline. When trained on large-scale simulated datasets then evaluated on real dataset, AAETR exhibits remarkable zero-shot sim-to-real transferability and emergent sidelobe suppression capability. This highlights the effectiveness of the proposed approach and its potential as a drop-in module in practical systems.

Figures (8)

• Figure 1: Zoom-in views of the BEV feature maps from AAETR, IAA and the radar detections by embedded DSP. AAETR demonstrates significantly lower FP detections and improved side-lobe suppression compared to the classic IAA method in the scene captured by the camera view in Figure \ref{['fig: front left camera view']}. Both AAETR and IAA provide much richer semantic information for downstream perception tasks than the sparse point clouds from on-device radars. For instance, the radar images from AAETR and IAA accurately capture vehicle dimensions in the scene and clearly detect thin street poles (highlighted by red circles). These low-level representations preserve valuable semantic information such as size and shape for subsequent perception tasks.
• Figure 2: The camera view of the front left side of the ego vehicle in the scene in Figure \ref{['fig: sample rendering 1']}.
• Figure 3: The overall network architecture. The radar signal and the array positions are fed into the transformer encoder as input. The encoded messages are then attended by the learnable queries in the decoder. The decoded messages are then mapped into the regression parameters and detection logits with FFN.
• Figure 4: The comparison of PR curves from AAETR and IAA when SNR is set to 35 dB. The precision-recall curves reveal how well each algorithm maintains detection performance across varying numbers of targets and dynamic ranges. The AATER generally shows better precision and recall trade off across wide range of threshold values.
• Figure 5: The comparison of the max F1 score and the mean error in magnitude and angle between AAETR and IAA across the number of targets per RD bin. The maximum F1 highlights the scalability of the AATER algorithm. The AATER maintains a significantly higher F1 score across all target numbers and SNRs. AAETR also generates more accurate magnitude prediction compared to IAA. Both methods are able to perform angle prediction with high accuracy. Overall, the robustness of the AAETR to varying levels of SNR is noticeably better than IAA.