An improved MADDPG-based framework for bidirectional safety management of construction robot interaction

Purpose Safety risks are particularly pronounced when multiple construction robots collaborate in complex construction environments. To enhance collaborative efficiency while ensuring operational safety, this study develops and evaluates, in a scaled simulation, a bidirectional safety management approach that balances autonomous operation, supervisory intervention, and obstacle-aware risk control. Design/methodology/approach A bidirectional multi-agent reinforcement learning framework is constructed for one operational robot (OPR) and one managerial robot (MAR). The proposed improved multi-agent deep deterministic policy gradient (improved-MADDPG) integrates reward shaping, curriculum learning and action constraints with controlled Gaussian exploration. A scaled cable-truss scenario is mapped into two-dimensional obstacle-free, sparse-obstacle and dense-obstacle environments for comparative and progressive ablation validation. Findings Within the evaluated simulation settings, improved-MADDPG improves convergence stability, target-reaching reliability, target-approach accuracy, and collision-risk suppression compared with standard MADDPG. Progressive ablation results further indicate that reward shaping, curriculum learning, and action constraints with controlled Gaussian exploration make complementary contributions under different obstacle complexities. Research limitations/implications The study advances safety-aware MARL for construction robotics by clarifying the cooperative-constrained OPR-MAR game and verifying the complementary effects of reward shaping, curriculum learning, and constrained MAR actions. However, validation remains limited to a 2D scaled simulation with one OPR and one MAR. Future research should test physical robots, incorporate multimodal perception and BIM/digital twin data, and evaluate multi-robot scalability, communication delay and critic complexity for larger robot teams. Practical implications The proposed method provides a procedural route for deploying supervisory construction robots that can monitor operational robots, maintain a safe separation distance and suppress collision-prone behaviour. The parameter table, scale mapping and ablation evidence improve reproducibility and help practitioners adapt the method to different construction stages. Integration with BIM/digital twin platforms could support real-time site-state updates, safety-zone visualisation, and robot trajectory supervision in congested workspaces. Social implications By improving robot safety management in complex construction environments, the proposed approach can help reduce worker exposure to collision hazards, unstable robot motion and high-risk manual inspection tasks. More reliable robot supervision may support safer human-robot coexistence, improve public confidence in construction automation and contribute to safer and more sustainable construction practices. Broader deployment should still consider workforce training, accountability, and regulatory compliance. Originality/value The study clarifies a supervisory game formulation for multi-construction robot safety management and demonstrates how safety-aware reward decomposition, progressive adversarial training and constrained MAR actions can be synergistically combined to support real-time bidirectional safety control in complex construction environments.