<p>Mixed traffic, where autonomous vehicles (AVs) share roads with human-driven vehicles, represents a long-term essential phase in the development of intelligent transportation systems. On-ramp merging of AVs in mixed traffic presents a bottleneck challenge due to complex non-cooperative multi-vehicle interactions. To address this issue, this paper proposes a novel game-theoretic motion planning approach termed the Multi-Level Cognitive Game Controller (MCGC). The MCGC integrates decision-making and control into a unified framework comprising four core modules: a trajectory generator, a decision-making mechanism, a belief estimator, and a motion planner. The trajectory generator produces trajectories as candidate vehicle actions by solving multiple tractable optimal control problems. The vehicle interactive decision-making mechanism is explicitly modeled as a partially observable multi-level cognitive game. A modified interacting multiple model is developed to estimate vehicle intentions online based on state observations. Finally, a stochastic model predictive control incorporating a two-stage game strategy is formulated to handle multi-vehicle interaction uncertainties, yielding an expected optimal solution while providing formal probabilistic safety guarantees. Simulation tests are conducted on diverse naturalistic datasets to validate the MCGC’s effectiveness, achieving an overall success rate of 97.9%. Simulation case studies further demonstrate the MCGC’s superior robustness compared to baseline game-theoretic controllers.</p>