Autonomous Magnetic Control of Capsule Robots with Deep Reinforcement Learning

Magnetically controlled capsule robots are valuable tools for gastrointestinal endoscopy, holding substantial clinical promise. To enable capsule robots to perform lesion detection, research on their motion control is essential. However, existing systems still rely on assistance from professionals to achieve full-process lesion detection and cannot yet achieve intelligent and autonomous motion control. To enhance autonomy in control, this paper presents a magnetically controlled capsule robot system utilizing deep reinforcement learning for autonomous operation. The system is equipped with visual sensors, magnetic sensors, and inertial sensors. Through a feature-level fusion strategy, multimodal environmental perception information is mapped into a unified state space. Within a deep reinforcement learning framework, the agent is trained to output actions to the robotic arm magnetic drive system, thereby achieving autonomous control of the capsule robot. Experiments on active motion control were conducted in environment of bend shape, and the results validate the effectiveness of autonomous control in the magnetically controlled capsule robot system.