Learning-based control for tendon-driven continuum robotic arms

Tendon-Driven Continuum Robots are widely recognized for their flexibility and adaptability in constrained environments, making them invaluable for most applications, such as medical surgery, industrial tasks, and so on. However, the inherent uncertainties and highly nonlinear dynamics of these manipulators pose significant challenges for classical model-based controllers. Addressing these challenges necessitates the development of advanced control strategies capable of adapting to diverse operational scenarios. This paper presents a centralized position control strategy using Deep Reinforcement Learning, with a particular focus on the Sim-to-Real transfer of control policies. The proposed method employs a customized Modified Transpose Jacobian control strategy for continuum arms, where its parameters are optimally tuned using the Deep Deterministic Policy Gradient algorithm. By integrating an optimal adaptive gain-tuning regulation, the research aims to develop a model-free controller that achieves superior performance compared to ideal model-based strategies. Both simulations and real-world experiments demonstrate that the proposed controller significantly enhances the trajectory-tracking performance of continuum manipulators. The proposed controller achieves robustness across various initial conditions and trajectories, making it a promising candidate for general-purpose applications.