Active Alignments of Lens Systems with Reinforcement Learning

TL;DR

This work tackles the challenge of aligning a multi-lens system to an imager under manufacturing tolerances. It casts active optical alignment as a POMDP and learns optimal alignment purely from pixel-space sensor observations using reinforcement learning (PPO) within the open-source relign simulation framework built on Mitsuba3. The study demonstrates that RL achieves faster, more robust alignments than Bayesian optimization and random baselines while maintaining high accuracy, even with noise in movements and lens tolerances. The relign framework, validated against optical design software and real alignment data, provides a realistic sim-to-real benchmark for high-precision optical assembly and enables efficient experimentation with minimal hand-crafted features. The results suggest significant practical impact for automated manufacturing and adaptive optics in reducing manual intervention and speeding up production.

Abstract

Aligning a lens system relative to an imager is a critical challenge in camera manufacturing. While optimal alignment can be mathematically computed under ideal conditions, real-world deviations caused by manufacturing tolerances often render this approach impractical. Measuring these tolerances can be costly or even infeasible, and neglecting them may result in suboptimal alignments. We propose a reinforcement learning (RL) approach that learns exclusively in the pixel space of the sensor output, eliminating the need to develop expert-designed alignment concepts. We conduct an extensive benchmark study and show that our approach surpasses other methods in speed, precision, and robustness. We further introduce relign, a realistic, freely explorable, open-source simulation utilizing physically based rendering that models optical systems with non-deterministic manufacturing tolerances and noise in robotic alignment movement. It provides an interface to popular machine learning frameworks, enabling seamless experimentation and development. Our work highlights the potential of RL in a manufacturing environment to enhance efficiency of optical alignments while minimizing the need for manual intervention.

Figures (8)

• Figure 1: Schematic presentation of a single alignment step, where a lens system consisting of four single lenses (left) has to be positioned relative to an optical sensor (right).
• Figure 2: Visualization of the two-dimensional projection $(s_i, s_j)\mapsto \|O(s^*-s_ie_i-s_je_j, W^*)-O^*\|$ for each tuple $\{i,j\}\subset\{x,y,z,R_x,R_y\}$, that is, all except two dimensions are fixed to the optimal values.
• Figure 3: Optical diagram of lens system including light rays in Code V.
• Figure 4: Comparison of point matching on a chessboard grid: relign (ours) vs. Code V.
• Figure 5: Comparison of a realistic alignment target: relign (ours) vs. a real alignment station.