Agricultural machinery operates under variable loads, impacts, dust, and changing soil–crop interactions, allowing faults to propagate through energy, material, information, and control pathways. This qualitative review synthesizes 195 research publications across fault formation, trustworthy diagnosis, prognostics and proactive risk control, maintenance and recovery, and case-based evidence assessment. Empirical findings are reported separately from review-derived recommendations, the authors’ conceptual requirements, and mandatory provisions of applicable standards or law. The literature most consistently supports controlled-fault identification, selected single-machine field monitoring, and localized operational compensation. Evidence is weaker for transfer across machines and seasons, calibrated prognostics, safety authorization, post-repair verification, and fleet-scale deployment. On this basis, the review recommends mission profile-specific fault boundaries, traceable diagnostic outputs, explicit uncertainty and abstention, and risk decisions linked to remaining work and resources. It also proposes a diagnostic passport, risk evolution trajectory, repair-effectiveness label, case-evidence matrix, and closed-loop repair workflow as review-level organizing constructs. No included study continuously followed the same machine or fleet through diagnosis, prognosis, authorization, maintenance, acceptance, and feedback; the framework therefore connects evidence-supported stages theoretically rather than claiming end-to-end validation.

High-value greenhouse crops require quality-sensitive selective harvesting under labor shortages and variable crop conditions. This review synthesizes 188 unique primary studies and examines robotic harvesting as a complete task chain rather than a set of isolated sensing or manipulation modules. Embodied intelligence is used as an analytical lens to connect perception, harvestability assessment, decision making, manipulation, feedback, failure propagation, and recovery. A greenhouse-cucumber case study illustrates how errors propagate across these stages. The synthesis shows that the dominant bottleneck is system integration: incomplete observability, short-lived scene states, contact uncertainty, and weak outcome verification or recovery cause cumulative losses in complete-task success, effective throughput, and product quality. Most reported systems remain at closed-loop automation or early embodied interaction, with limited evidence of interaction-driven learning or cross-context generalization. We therefore frame evaluation around six production-relevant dimensions: complete-task success, effective cycle time and retry cost, product quality, failure detection and recovery, human intervention, and continuous operation. Near-term priorities are reliable, recoverable operation in defined crop–facility systems and standardized reporting; adaptation across cultivars and greenhouse configurations is a medium-term objective, whereas embodied learning and broad cross-platform transfer remain longer-term research directions. This system-level perspective reframes smart robotic harvesting around information continuity, fault containment, and deployable greenhouse performance.