MILD: Mediator Agent System for Enhanced Human-Vehicle Collaboration
Researchers have introduced the Mediator-in-the-Loop-Driving (MILD) system, a novel agentic architecture designed to improve human-vehicle collaboration in partially automated driving. Addressing the cognitive burden caused by bidirectional misalignment between drivers and automated systems, MILD elevates the human role from passive supervisor to active manager. The system integrates a perception agent for comprehensive in-cabin and out-of-cabin understanding with a lightweight strategy agent that generates explainable action suggestions. To ensure safety and alignment with human values, the developers created Evidence- and Constraint-weighted Policy Optimization (ECPO), which uses automatic validators to prevent constraint violations. Additionally, a retrieval-augmented generation module dynamically incorporates traffic regulations and driver preferences into decision-making. Field experiments across three open datasets demonstrate that MILD outperforms baseline models in perception accuracy, strategy quality, and human-rated comfort and explanation adequacy. This work provides a practical pathway for building auditable, aligned agents for safer and more intuitive collaborative driving experiences, marking a significant advancement in automotive artificial intelligence.
Wire timeline
MILD: Mediator Agent System for Enhanced Human-Vehicle Collaboration
Researchers have introduced the Mediator-in-the-Loop-Driving (MILD) system, a novel agentic architecture designed to improve human-vehicle collaboration in partially automated driving. Addressing the cognitive burden caused by bidirectional misalignment between drivers and automated systems, MILD elevates the human role from passive supervisor to active manager. The system integrates a perception agent for comprehensive in-cabin and out-of-cabin understanding with a lightweight strategy agent that generates explainable action suggestions. To ensure safety and alignment with human values, the developers created Evidence- and Constraint-weighted Policy Optimization (ECPO), which uses automatic validators to prevent constraint violations. Additionally, a retrieval-augmented generation module dynamically incorporates traffic regulations and driver preferences into decision-making. Field experiments across three open datasets demonstrate that MILD outperforms baseline models in perception accuracy, strategy quality, and human-rated comfort and explanation adequacy. This work provides a practical pathway for building auditable, aligned agents for safer and more intuitive collaborative driving experiences, marking a significant advancement in automotive artificial intelligence.
cs.AI updates on arXiv.org