Review of Key Technologies for Harvesting Embodied Intelligent Robots in Facility Horticulture: From Multimodal Active Sensing to Self-Evolving Closed Loops

Abstract

Embodied intelligence offers a system-level paradigm for improving the adaptability of robotic harvesting in complex horticultural environments by coupling perception, decision-making, and physical interaction. This review systematically analyzes recent advances in embodied intelligence for robotic harvesting. A search of the Web of Science Core Collection covering January 2020 to August 2026 identified 294 records, of which 100 studies were included after screening. Based on the reviewed literature, a literature-grounded Four-Tier framework is proposed, comprising Multimodal Active Sensing, Semantic Context Recognition and Understanding, Knowledge-Driven Decision-Making and Experience Evolution, and Skilled and Compliant Execution. The review shows that embodied approaches have improved individual capabilities across perception, decision-making, and manipulation; for example, representative studies reported over 12% improvement in detection accuracy, reduction in fruit damage from 8.2% to 2.0%, and 89% retention in continual learning. However, reliable autonomous harvesting remains constrained by cross-module uncertainty, real-time crop–robot interaction, continual adaptation, and the lack of standardized benchmarks. Future research should emphasize physically grounded multimodal closed-loop integration, experience-driven adaptation, standardized evaluation, and crop–robot co-design.