Machine learning (ML) calibration techniques could be an effective way to improve the performance of low-cost air pollution sensors. Here we provide novel insights from case studies within the urban area of London, UK, where we compared the performance of three ML techniques to calibrate low-cost measurements of NO₂ and PM₁₀. In particular, we highlight the key issue of the method-dependent robustness in maintaining calibration skill after transferring sensors to different measurement sites.