Towards Real-world Lens Active Alignment with Unlabeled Data via Domain Adaptation

TL;DR

This work tackles the domain gap hindering real-world deployment of active alignment (AA) models trained on simulated optical data. It introduces Domain Adaptive Active Alignment (DA3), combining tolerance-aware simulation with a minimal set of unlabeled real-world images, a domain transformation generator (autoregressive, G) to produce pseudo-target data, and domain-adaptive training that enforces domain-invariant degradation features through degradation-based augmentation and feature alignment. Empirical results on two lens setups show that DA3 significantly reduces misalignment error, reaching MAE around $2.03~\mu$m and approaching dense on-device labeling performance while drastically cutting data-collection time. The approach validates a practical digital-twin pipeline for scalable, cost-effective large-scale optical assembly by bridging simulation and real-world imaging via domain adaptation.

Abstract

Active Alignment (AA) is a key technology for the large-scale automated assembly of high-precision optical systems. Compared with labor-intensive per-model on-device calibration, a digital-twin pipeline built on optical simulation offers a substantial advantage in generating large-scale labeled data. However, complex imaging conditions induce a domain gap between simulation and real-world images, limiting the generalization of simulation-trained models. To address this, we propose augmenting a simulation baseline with minimal unlabeled real-world images captured at random misalignment positions, mitigating the gap from a domain adaptation perspective. We introduce Domain Adaptive Active Alignment (DA3), which utilizes an autoregressive domain transformation generator and an adversarial-based feature alignment strategy to distill real-world domain information via self-supervised learning. This enables the extraction of domain-invariant image degradation features to facilitate robust misalignment prediction. Experiments on two lens types reveal that DA3 improves accuracy by 46% over a purely simulation pipeline. Notably, it approaches the performance achieved with precisely labeled real-world data collected on 3 lens samples, while reducing on-device data collection time by 98.7%. The results demonstrate that domain adaptation effectively endows simulation-trained models with robust real-world performance, validating the digital-twin pipeline as a practical solution to significantly enhance the efficiency of large-scale optical assembly.

Figures (8)

• Figure 1: Comparison of optical active alignment learning paradigms. (a) Traditional supervised learning yields high accuracy but requires high costs for on-device data collection. (b) Pure simulation-based learning is free of on-device data collection yet suffers from low accuracy due to significant domain gaps. (c) Our proposed framework combines pure simulation with minimal unlabeled real-world data through domain adaptation, thereby achieving high accuracy at a very low on-device data cost.
• Figure 2: Trade-off analysis between mean prediction error and data collection time cost. The dashed curve illustrates the efficiency frontier of conventional supervised methods, highlighting the trade-off between accuracy and data cost. On-device ($N_{oracle}$) denotes training on data collected from $N_{oracle}$ real-world lenses. *Real-world data collected with sparse sampling. $^{\dagger}$Simulation data without tolerance perturbation.
• Figure 3: Illustration of the domain gap at representative decenter offsets (Unit: µm). Although the simulation images (top) and the on-device images (bottom) exhibit similar degradations due to the same misalignment, there are distinct differences in their imaging styles.
• Figure 4: Overview of the proposed domain adaptive active alignment (DA3) framework. (a) The source dataset is constructed by grid sampling in a calibrated misalignment space through tolerance-aware optical simulation, while the target dataset is collected from several unlabeled random misalignment positions on a real-world AA machine. (b) The domain transformation module bridges the domain gap by transferring real-world imaging styles to simulated images, thereby generating content-consistent paired training data. (c) Domain adaptive training strategies achieve feature alignment across different domains by adding pixel-level consistency constraints to features extracted from degradation-based augmented paired data and performing adversarial training on them.
• Figure 5: Optical structures of the experimental lenses and the real-world AA machine. (a) The black-box model of the security lens. (b) The 7P aspheric smartphone lens. In both diagrams, the lens groups enclosed in dashed rectangles indicate the movable components adjusted during the AA process, while the remaining elements are fixed. (c) The real-world AA machine setup utilized for on-device data acquisition and lens assembly.