Human-Agent Shared-Control Based Training Scheme for Efficient Robotic Tadpole Navigation

Real-world deployment of reinforcement learning (RL) in multijoint, multimode robotic systems remain challenging due to underactuated dynamics, limited onboard computational resources, sparse and task-specific reward design, unstable fluid-structure interactions, and frequent long-tail failures. We propose a human-agent shared control framework that couples an RL policy with risk-aware arbitration, transferring control to a human when policy uncertainty indicates elevated task risk. Human demonstration is internalized via four techniques: prioritized demonstration-driven experience replay, intervention-based reward shaping, human-guided learning objective, and uncertainty-guided authority switching. Experiments in simulation and on hardware validate improved training efficiency over advanced RL strategies, strong generalization across diverse actuation parameters, trajectory complexities, and environmental disturbances, as well as reliable sim-to-real transfer. Furthermore, our framework allows brief, targeted human intervention for real-time fine-tuning of posttrained agents. These results indicate that the human-agent shared control scheme offers a practical, data-efficient solution for robust RL deployment in dynamic, unstructured environment of such real robotic systems. © 2026 IEEE.