Multi-Target Harvesting Robots: Key Advances and Future Trends in Perception, Decision-Making, and Task Scheduling

Multi-target harvesting is an important direction for improving the efficiency of orchard harvesting robots and advancing their engineering deployment. Compared with single-target harvesting, multi-target harvesting is constrained by dense fruit distribution, branch and leaf occlusion, illumination variation, manipulator accessibility, harvesting-sequence planning, and multi-agent conflicts. These factors impose higher requirements on visual perception, state representation, decision-making, and task scheduling. This paper focuses on multi-target harvesting tasks in complex orchard environments. It first analyzes task characteristics and constraint sources and then reviews advances in visual perception and scene understanding for target recognition, attribute identification, spatial localization, and spatiotemporal consistency maintenance. Typical decision-making methods are summarized, including rule- and experience-based heuristics, cost-model-based optimization, and data-driven learning approaches. Progress in multi-arm and multi-robot harvesting systems is further reviewed in terms of perception sharing, task allocation, cooperative operation, and conflict avoidance. Continuous and stable perception under complex occlusion, unified scheduling-oriented state representation, real-time decision-making under strong constraints, collision avoidance in shared multi-agent workspaces, and distributed state synchronization remain key challenges. Overall, multi-target harvesting has shown potential to improve orchard operation efficiency, but efficient collaboration and engineering deployment in complex orchard environments still require further breakthroughs.