A simulation study of decomposition-based inverse reinforcement learning toward construction automation: The case of excavator in challenging environments

The increasing demand for construction in unstructured and dynamic field conditions requires robots capable of autonomous operation. While imitation learning has shown promise, existing approaches are inherently resource-inefficient and primarily imitate human behaviors rather than underlying decision-making strategies, limiting performance in such conditions. This paper proposes an inverse reinforcement learning-based framework to address these challenges. The framework decomposes construction tasks into subtasks with explicit state features to structure the robot’s decision-making, infers model parameters from limited expert demonstrations, and applies them to the architecture to enable task execution. Simulation-based experimental results demonstrate that the agent trained with the proposed framework achieved a 73% task success rate, outperforming agents trained with traditional approaches. Furthermore, the experiments demonstrated the capacity of IRL to reproduce strategic decision-making of human experts under dynamic and unstructured construction conditions. This study enhances the autonomy of construction robots, further contributing to automation in construction. • A modular inverse reinforcement learning for construction robotics • Decomposes tasks into subtasks and explicit state features in extreme construction • Integrates physical and functional improvements for realistic simulations • Expert demonstrations and parameter estimation to enhance the generalization • Achieves 73% task success rate in dynamic and complex work settings