A teleoperation-guided incremental preference learning approach for wheel mobile robots

Reinforcement-learning-based navigation for wheeled mobile robots is commonly optimized using environment-defined rewards. Although such policies can achieve goal-reaching behaviors, they may exhibit limited generalization in untrained scenarios and fail to reflect human safety preferences. To address these issues, this paper proposes a teleoperation-guided incremental preference learning framework for WMR cooperative navigation. During the navigation process of the WMR, intermittent teleoperation interventions are introduced only when the autonomous policy produces unsafe or non-preferred behaviors. The teleoperated trajectories are paired with the corresponding policy-generated trajectories to construct preference comparisons, from which a BiGRU-MLP preference network learns implicit human path preferences and generates preference rewards. Furthermore, an environment-preference decoupled actor-critic architecture is developed. The critic network is trained using environment rewards to preserve basic goal-reaching and obstacle-avoidance capabilities, while the actor network is incrementally updated using mixed advantages based on preference rewards to align the policy with human-preferred behaviors. Experiments demonstrate that the proposed method achieves a higher success rate than pure autonomous navigation and improves efficiency, increasing the average speed by 68.5% over teleoperation-based shared control, while also enlarging the safety margin to 0.56 m and reducing the frequency of human interventions..