Prediction-Enhanced Artificial Potential Field Control for Collision Avoidance in Human–Robot Collaboration

Safe human–robot collaboration requires a manipulator to respond to human motion without repeatedly interrupting the task. Traditional artificial potential field (APF) control is computationally efficient, but it can become trapped in local minima and may react too late to a rapidly moving obstacle. This paper presents a prediction-enhanced APF framework for a UR5 manipulator. The controller reconstructs attractive acceleration to retain goal guidance near spherical and capsule-shaped obstacles and represents a 0.5 s hand-position forecast as a virtual obstacle that generates an auxiliary repulsive field. Future hand positions are predicted from upper-limb joint sequences using a graph convolutional network–bidirectional long short-term memory (GCN-BiLSTM) encoder and an attention-based autoregressive decoder trained with scheduled sampling–curriculum learning. On 31 held-out hand-trajectory files, the predictor achieved a mean frame-15 error of 0.02942 m, with 89.52% of test windows within 0.05 m. In simulation, the modified APF avoided stagnation in the tested static configuration. Across 31 dynamic trajectory simulations, model prediction increased avoidance success from 16.1% without prediction to 96.8%. Mean processing latency from hand-frame processing to scene-marker observation was 19.03 ms. These results indicate that short-horizon hand forecasts improved local collision avoidance under the tested simulation conditions.