Domain-Model-Map-based reinforcement learning for feedforward anti-sway control of overhead cranes

다국어 초록 (Multilingual Abstract)

Overhead cranes are widely used in industries such as shipbuilding and construction, enabling efficient handling of heavy loads. However, achieving precise control of overhead cranes remains a challenge, primarily due to undesired oscillations of the heavy loads, especially under varying rope lengths. Traditional feedforward control methods are effective only under constant rope length, limiting performance during simultaneous hoist and trolley operation. Feedback-based approaches can compensate for changing conditions but often rely on complex sensor integration and continuous velocity adjustments, which may increase system complexity and operation cycle time. To address these practical challenges, this study proposes a reinforcement learning (RL)- based feedforward control approach that integrates a model-based offline ensemble with online selection to maintain high performance across diverse conditions. In a simulation crane model environment, the RL agent is trained to generate trajectories that minimize both residual sway angle and operating time. The proposed strategy pre-identifies optimal models for specific operating domains, allowing for efficient execution by selecting a single optimal model during the online stage. Although this approach involves a trade-off requiring substantial offline training to construct the model map, it reduces online complexity suitable for industrial PLCs.
This approach minimizes the real-time computational burden compared to conventional ensemble methods, generating near-optimal trajectories under varying rope lengths without relying on explicit sensor feedback. We validated this method through simulations and experimental tests, demonstrating a 38% reduction in operation time and effective sway suppression with residual angles under 0.3◦, which demonstrates its feasibility for practical industrial applicability.