6G networks are envisioned to enable the Internet of Things (IoTs) and foster ubiquitous sensing. Urban opportunistic crowdsensing, which leverages participants carrying mobile sensing units (MSUs) in their daily activities to collect data, enables low-cost and large-scale urban sensing for applications such as air quality monitoring, pothole detection, and noise classification. However, urban opportunistic crowdsensing poses challenges to conventional cloud-based centralized learning due to the unpredictable nature of crowdsensed data collection and privacy concerns from uploading personal information to the centralized cloud. Federated learning (FL), where MSUs act as data sources and computing nodes, offers a promising alternative to mitigating these issues. Despite FL’s potential, urban crowdsensing contexts’ spatial-temporal diversity, mobility, constrained resources, and emerging privacy concerns present new challenges. In this paper, we explore the opportunities and challenges of FL in urban opportunistic sensing in 6G networks and suggest potential optimization strategies. Furthermore, we conduct field experiments in Helsinki, Finland, and design an FL-based air quality calibration method for opportunistic crowdsensing to demonstrate the feasibility of our vision.