UNA-SAC: An Uncertainty-Aware Nonlinear Arbitration Method for Human–AI Shared Control

With the continuous development of artificial intelligence (AI), human-AI shared control has become an essential paradigm for achieving reliable collaboration, where the key challenge lies in efficiently arbitrating between human and AI policies. However, the inherent uncertainty of AI policies and their approximation errors often undermine the robustness and effectiveness of traditional linear arbitration. To address this issue, this paper proposes a nonlinear arbitration method based on the Soft Actor-Critic (SAC) framework, termed UNA-SAC. The method introduces a moment network to model AI policy uncertainty and incorporates a cognition-inspired mechanism to adjust the human policy, thereby constructing a distributional nonlinear arbitration form. Theoretical analysis demonstrates that the proposed method provides advantages in gradient optimization and effectively mitigates the cumulative effect of uncertainty-induced bias. Experimental results further validate its superiority in driving assistance scenarios: UNA-SAC achieves significant improvements in convergence speed, task success rate, robustness, and operational performance compared with linear arbitration and other baseline methods.