Research Brief: How AI-guided Greenhouses Can Maximize Profits
April 16, 2025 Computer scientists at the University of California, Davis (UC Davis) explored the challenges and opportunities of using fully automated artificial intelligence (AI) systems in greenhouses in a recent study.
The study, led by Yongshuai Liu, Taeyeong Choi, and Xin Liu, outlined a framework for autonomous greenhouse management that combines reinforcement learning (RL) and human decision-making. Their work highlights how policy learning – a concept commonly applied in robotics and gaming– can be used in agriculture.
Study Recap
In greenhouse environments, AI makes important decisions based on temperature, lighting, irrigation, and CO2 levels to increase crop yield while minimizing resource use. In the study, researchers defined common objectives growers have, such as maximizing net profit, balancing energy costs, and adhering to safety limits, and how AI could better assist them.
Since AI works best with fast-evolving datasets and plant growth is often slow with little data attached to it, agriculture presents several roadblocks for AI. To address these challenges, the researchers created predictive models that simulated greenhouse environments. This allowed AI to make decisions without physical trial and error and learn faster from the outcomes. While their results were promising, the researchers noted that AI’s accuracy started to decline during time-consuming agricultural tasks.
The researchers also noted system-level challenges in the study. These include the simulation-to-reality gap, where control strategies trained on virtual models don’t perform well in real-world conditions. To bridge this gap, researchers recommend integrating a domain adaptation that helps transfer the policies AI learned to real greenhouse environments. They also stressed the need for AI to justify its decisions in ways growers can understand and trust, noting that human-AI interaction is essential.
Looking Ahead
Despite their positive findings, the researchers warned that fully autonomous greenhouse control will continue to evolve for the foreseeable future. They emphasized the need for deeper investment in explainable and adaptive AI systems that can handle long-term planning, real-time decision-making, and the variability of agriculture. For the full research study, go here: https://arxiv.org/html/2503.21640v1.