Introduction
In many undergraduate curricula, autonomous systems are synonymous with mobile robotics. While foundational, these courses often spend significant time on hardware-specific challenges like kinematics and localization, potentially overshadowing "higher-level" AI integration challenges. Researchers at Carnegie Mellon University addressed this gap by developing a course focused on the deployment of autonomous agents within a stationary but complex environment: the autonomous greenhouse.
The TerraBot Platform
The centerpiece of the course is the "TerraBot," a 10-gallon fish tank converted into an automated garden. This platform provides a self-contained system for studying perception, decision-making, and long-term autonomy.
The hardware includes:
- Perception: Redundant sensors for temperature, humidity, light, and soil moisture to account for sensor noise, alongside a color camera and ultrasonic water level sensors.
- Actuation: Controllable red/blue LED grow lights, a water pump, and fans.
- Computation: An Arduino handles low-level control, while a Raspberry Pi running the Robot Operating System (ROS) manages high-level AI logic.
Methodology and Simulation
The curriculum is scaffolded across seven units, progressing from basic ROS infrastructure to advanced topics like Kalman filters for modeling uncertainty and computer vision for plant health monitoring. A significant challenge in botanical robotics is the slow growth cycle of plants. To mitigate this, the researchers developed a 3D graphical simulator.
This simulator allows students to "speed up time," testing their agents' logic against simulated evaporation, transpiration, and fluctuating ambient light before deploying them for two-week "Grow Periods" on real hardware.
Key Insights: Stationary Complexity
The course demonstrates that a stationary system can be as challenging as a mobile one. Students must navigate noisy sensors and potential actuator failures (e.g., a pump failing to turn off). A notable pedagogical success was the "email task," which required agents to send daily status reports. This forced students to implement explainable AI, as the agent had to justify its actions (e.g., "I activated the pump because soil moisture fell below the 20% threshold"), which proved essential for debugging long-term autonomous behavior.
Results and Discussion
Since its inception in 2019, the course has seen enrollment nearly triple, with high student satisfaction ratings (4.8/5). Technically, the majority of student teams successfully grew radishes to full maturity (approximately 19 days) using entirely autonomous agents. However, growing lettuce remained a "hard mode" challenge, often requiring more precise climate control than many initial algorithms provided. Future iterations of the course aim to incorporate Deep Reinforcement Learning and provide more iterative growth periods to refine student control strategies.
Summary updated