RTSP-Net: RL-Based Robotic Task Sequencing for Robotic Manipulators in Obstacle-rich Environments 

Abstract: The Robotic Task Sequencing Problem (RTSP) for manipulators involves determining the order in which targets are visited and selecting a joint configuration for each target pose from multiple inverse-kinematics solutions. Existing approaches often use straight-line distance approximations to estimate motion costs, which can underestimate obstacle-induced detours in cluttered environments. While precomputing pairwise motion costs is practical for fixed targets in static settings, it becomes computationally expensive in applications such as inspection, cleaning, and disinfection, where target sets and environments may change. To address this, we propose RTSP-Net, a hierarchical reinforcement learning framework that takes a scene point cloud, target poses, and their candidate manipulator configurations as input and outputs a visiting sequence with one configuration per target. RTSP-Net is trained across six environment types with varying obstacle layouts, using a scene-conditioned neural cost model to estimate transition costs and provide training rewards. Experiments show sequence-length reductions of up to 27% in in- distribution environments and 19% in out-of-distribution scenarios without retraining. RTSP-Net maintains sub-second inference for all tested problem sizes up to 500 targets, achieving up to 43× speedup (using GPU acceleration) in configuration selection and sequence generation over the baseline.