Persistent Autonomy for Phenotyping: A Reconfigurable Multi-Robot Inspection System with Online Task Reallocation and Visual Policy Adaptation

High-throughput plant phenotyping demands persistent, autonomous crop inspection across large fields over days or weeks, yet existing agricultural robot systems assume static mission plans that cannot react to changing crop conditions, weather disruptions, or robot failures. We present a reconfigurable multi-robot inspection system that continuously monitors field plots, dynamically reallocates sensing tasks among heterogeneous ground and aerial platforms, and adapts its visual inspection policies online as crop appearance changes through growth stages. The system comprises three integrated subsystems: (i) a mission persistence layer that maintains long-horizon task queues with priority decay, interrupted-task resumption, and weather-aware scheduling, (ii) a market-based task reallocation engine that redistributes inspection waypoints among available robots when platforms join, leave, or degrade, and (iii) an online visual policy adaptation module that specializes pretrained crop inspection models to current field conditions using a lightweight experience replay mechanism. We deploy the full system on a fleet of two Clearpath Jackal UGVs and one DJI Matrice 300 UAV inspecting a 4-hectare sorghum trial at the University of Nebraska–Lincoln research farm over a 21-day campaign. Compared to a fixed-schedule baseline, our system increases plot coverage by 34%, reduces revisit latency by 41%, and maintains inspection accuracy above 91% despite substantial canopy appearance change across growth stages.