Potential Fields for Position-Constrained Autonomous Motion in Robotic Imitation Learning

Bilateral teleoperation has facilitated imitation learning that incorporates force control. However, conventional imitation learning methods lack an obstacle avoidance mechanism, making them unable to adapt to environmental changes. Consequently, whenever the environment is altered, data must be recollected and the entire learning process repeated. This study introduces a potential-field-based approach to enable obstacle avoidance in environments containing obstacles that were not present during the learning phase. Specifically, we propose an “orthogonal potential” method that applies a virtual force perpendicular to the robot's velocity. This allows the robot to steer smoothly around obstacles without the deceleration often caused by conventional repulsive fields. Nevertheless, learning models with recursive structures may exhibit unstable or erroneous behaviors due to their non-Markovian nature. To mitigate this issue, we employ a hierarchical model in which the lower layers preserve the Markov property, thereby enabling adaptive behavior within the potential field. After training via bilateral control-based imitation learning, we evaluated whether the proposed approach could adapt to environmental changes during autonomous pick-and-place tasks. Experimental results demonstrate that the proposed method improves the task completion rate compared to the conventional repulsive potential method, achieving robust avoidance even for obstacles placed directly on the trajectory.