Real-time adaptive incentive control via Stackelberg game for human–machine collaborative engineering systems

Abstract

This paper presents a real-time adaptive incentive-control framework for hybrid human–machine collaborative engineering systems, addressing human bounded rationality, cross-task policy transfer, and millisecond-level response under equilibrium constraints. An asymmetric two-level Stackelberg model combines prospect-theory value functions with quantal response to represent heterogeneous human and machine decision processes. An online meta-reinforcement learning algorithm enables rapid adaptation across tasks, while a fast-timescale event-triggered regulator provides practical convergence to a noise-dependent equilibrium neighborhood and excludes Zeno behavior. Simulation and hardware experiments with 24 human participants and four robotic agents cover assembly-line, warehouse-logistics, and emergency-rescue scenarios. The proposed method achieves task completion rates of 89.6%–92.7%, human compliance rates of 83.2%–87.4%, convergence in 450 episodes rather than 800 for MAML-RL, and zero-shot cross-scenario completion rates of 64.7%–68.9%. Event triggering reduces communication frequency from 200 to 54.3 Hz, with a 95th-percentile edge-inference latency of 27.1 ms. Ablation results support the separate contributions of prospect-based valuation, quantal response, and influence-factor correction. The evidence establishes prospective personalization for calibrated operators; transfer to previously unseen operators remains to be validated.