A reinforcement learning-based path-planning method for cable-driven hyper-redundant robots in unknown environments

Cable-driven hyper-redundant robots (CDHRRs) show great potential for confined-space detection due to their dexterity and adaptability. However, the redundant degrees of freedom (DoFs) present significant challenges for CDHRRs in planning feasible paths with stable configurations. To address this challenge, this paper presents a path-planning algorithm to improve the stability and applicability of generated paths. Firstly, the Soft Actor–Critic (SAC) algorithm is used to efficiently plan paths for CDHRR. To address the issue of unstable configurations, a path smoothness index and a reward function are designed to evaluate and enhance the feasibility of generated paths. Then, the Hindsight Experience Replay (HER) and Prioritized Experience Replay (PER) algorithms are applied to solve the sparse-reward problem of CDHRRs with corresponding scenarios. Finally, the proposed method is validated via both simulation and real-world experiments. Simulation results show that, compared to the Rapidly-exploring Random Tree Star (RRT*) and Arc-segment RRT (As-RRT), the computation time of the presented algorithm is reduced by 96.56% and 97.95%, respectively. Moreover, this deep reinforcement learning (DRL)-based algorithm generates smoother paths with a more stable configuration. In real-world experiments, the proposed algorithm demonstrates a 14.9% and 10% improvement in path smoothness and success rate over As-RRT. • An specialized SAC framework for large-scale CDHRRs in unknown environments. • Instability-aware reward design and path smoothness index for feasible configurations. • HER or PER integrated for sparse rewards with corresponding scenarios. • Validated in simulation and real-world, enhancing robustness and task success rate.