Safety-Aware Asymmetric Policy Architecture for Autonomous Agents in Hetero geneous Environments: AGrey-box Continuous Modulation Approach

Achieving a balance between safety, efficiency, and motion smoothness remains a critical challenge for autonomous agents operating in high-density heterogeneous multi-agent environments. We argue that this dilemma can be effectively addressed by recognizing the distinct operational nature of driving sub-tasks: the need for continuous, fine-grained regulation in longitudinal control versus the strategic decision making process involved in lateral moves. To address this, we propose a Grey-box Asymmetric Driving Architecture (GADA). This framework structurally decouples the policy into heterogeneous streams: a Continuous Longitudinal Behavior Modulation (CLBM) mechanism that adaptively smooths a baseline physics-based kinematic model for human-like motion quality, and a discrete lateral policy strictly guarded by an Explicit Safety Shield to ensure zero-collision feasibility. Extensive experiments in specific heterogeneous scenarios demonstrate that GADA comprehensively outperforms baselines. Specifically, compared to physics based benchmarks, the proposed architecture reduces average jerk by 72.5%, significantly enhancing motion smoothness while maintaining high system throughput and ensuring a 0.00\% collision rate. These results validate that the proposed grey-box paradigm offers a robust, neuro-symbolic solution for the cooperative integration of autonomous systems into complex human-populated environments.