This project integrates a reinforcement learning controller, camera-based payload yaw estimation, and real-time execution on Raspberry Pi to build an image-based payload yaw control system.

The results demonstrate not only that payload yaw can be controlled in simulation, but also that image-based feedback control can be executed on an edge device.

Potential for Real-World Applications

Although the control model was trained in a physics simulation, it uses the yaw angle estimated from camera images during control, rather than the ground-truth yaw provided by the physics engine.

By separating image-based state estimation from the control model, the same system architecture can be extended to real-world applications using camera images as input.

Real-Time Control on Raspberry Pi

Both image-based pose estimation and the control model were executed on Raspberry Pi while maintaining a 3 Hz control cycle.

The control performance on Raspberry Pi was comparable to that in the development environment, demonstrating that image-based feedback control can be implemented on a compact edge computer without a high-performance GPU.

Applications to Other Control Systems

This project focuses on payload yaw control using four thrusters.

However, the overall architecture—estimating the system state from camera images and using the estimated state as input to a control model—is not limited to thruster control. By adapting the training environment and control outputs to different actuators and target systems, the same approach can be applied to image-based feedback control for transport equipment, mobile robots, and attitude control systems.

Toward Real-World Deployment

This project verified the effectiveness of the proposed image-based control system in a simulation environment.

Applying the system to real hardware will require additional validation, including differences between simulated and real camera images, environmental disturbances, actuator dynamics, and output limitations. Addressing these factors will enable the system to move closer to practical real-world applications.