Navigation of Tendon-driven Flexible Robotic Endoscope through Deep Reinforcement Learning

Robotic endoscopes play a crucial role in diagnosing gastrointestinal disease and performing tumor resections. While current research primarily focuses on autonomously controlling rigid robots, establishing control models for flexible robots remains challenging. To address this, model-free deep reinforcement learning (DRL) presents a promising approach for enabling agents to make decisions under uncertainty. In this paper, we investigate the control policy of a flexible endoscope using Simulation Open Framework Architecture (SOFA) platform. We design a flexible tendon-driven robotic endoscope (TDRE) and develop a custom simulation environment within SOFA to train DRL agents. Our approach involves implementing the Proximal Policy Optimization (PPO) algorithm to approximate an optimal policy for trajectory planning. The optimal policy facilitates trajectory tracking tasks for the TDRE’s end-effector, such as circle trajectories and action disturbances, without requiring fine-tuning policy network parameters. Experimental results demonstrate that our approach achieves near real-time performance (30 FPS). The feedforward neural network of the policy provides feedback, enabling closed-loop control of TDRE. Furthermore, our experiments show that the navigation success rate of TDRE exceeds 90% within a tolerant error of 3 mm in free space. Notably, compared to direct training with contact, navigation tasks with contact retrained by a pre-trained policy in free space exhibit enhanced navigation capabilities.