Demonstration using Reinforcement Learning, Image-based Pose Estimation, and Raspberry Pi

Payload Yaw Control

This project develops an autonomous payload yaw control system
by combining image-based pose estimation with reinforcement learning.
The estimated yaw angle is used to control four thrusters in real time.

Part 2 — Latest

Low-Power Variable Target Yaw Control

A single control model supports target yaw angles of 0°, ±45°, and ±90°. Low-power control is evaluated using three payload types.

Overview Read Part 2

Part 1

Image-based Payload Yaw Control

A baseline system that estimates payload yaw from camera images and executes the learned controller in real time on Raspberry Pi.

Overview Read Part 1

Part 2 Overview

Part 2 extends the baseline system by introducing variable target yaw control and low-power thruster control. The controller is evaluated for three payload types at 0°, ±45°, and ±90°.

Part 1 Overview

Part 1 presents the baseline payload yaw control system combining image-based yaw estimation, reinforcement learning, and real-time execution on Raspberry Pi.

This demonstration shows real-time payload yaw control on Raspberry Pi using yaw angles estimated from camera images. The left shows the proposed control system, while the right shows the same conditions without control.