New Framework Audits RL Rewards for Smart Greenhouses

Reinforcement learning (RL) offers a powerful way to test climate control strategies in smart greenhouses, providing speed and scale that traditional crop experiments cannot match. However, simply getting a single simulator return is insufficient for growers and control engineers. They need to understand how the policy achieves its results, such as when it heats, enriches CO2, vents, or manages humidity. To address this, the research introduces a calibration-first reward audit framework. This framework systematically breaks down the overall scalar reward into its constituent components, including conditional terms for temperature, CO2, humidity, vapor-pressure-deficit, screen usage, and actuation proxies. This decomposition allows for consistent comparison of these reward components across various stages, from simulator training to facility-adapted rollouts and logged data from challenges like the Autonomous Greenhouse Challenge. By applying this framework within GreenLight-Gym, the researchers adapted the simulator to real-world climate traces and successfully scored the same reward components on logged greenhouse data. This approach provides greater transparency and interpretability for RL-driven climate control policies, enabling better understanding and refinement of smart greenhouse operations.