Dengbo He

Achieving safe and precise supervised autonomous landing of eVTOL (electric vertical take-off and landing) aircraft in environments lacking special infrastructure remains a critical challenge. Existing approaches either rely on specific cooperative visual markers (which perform poorly in unstructured settings) or on highly precise global navigation satellite system (GNSS) data and multiple sensors. Meanwhile, large vision-language models (VLMs) excel at high-level semantic reasoning yet lack accurate 3D perception and stable control – leading to a spatial grounding problem. Therefore, this study proposes Agent-Pilot, a modular hybrid-intelligence framework that enables supervised autonomous landing without requiring specialized infrastructure. We propose a purely monocular vision-based approach that integrates classical Proportional-Integral-Derivative (PID)-based feedback control, 3D spatial perception, real-time video-based depth estimation, and VLM-driven decision-making into a unified pipeline. The approach requires no pre-training on specific landing markers. Finally, we introduce a decoupled safety agent to monitor internal and external anomalies and alert the pilot in the event of an abnormality. The framework is evaluated in a flight simulation through online testing in Microsoft Flight Simulator (MSFS) 2024, realized through a self-developed toolbox, the MSFSDevKit, which is open-sourced in this study. Results demonstrate that Agent-Pilot achieves high success rates across 16 challenging scenarios and maintains attitude stability and glide-path accuracy superior to those of trained human pilots in both clear and fog conditions at the urban vertiport. All necessary code and tools used in this study are open-sourced to enable researchers to replace components modularly to improve the framework.

As Advanced Driver Assistance Systems (ADAS) become increasingly common in production vehicles, surrounding drivers may increasingly encounter vehicles whose behavior is partly controlled by automation. In China, some ADAS-supported vehicles use an externally visible indicator, commonly referred to as the "small blue light," to signal that ADAS is active. However, limited naturalistic evidence is available on whether such external ADAS status visibility influences surrounding drivers. This study investigated following drivers’ car-following behavior in naturalistic traffic when the lead vehicle operated with ADAS activated and the small blue light was on or off. Results showed speed-dependent changes under the blue-light-on condition. In particular, following drivers maintained larger time headway when the lead vehicle traveled at higher speeds with the light on. By providing naturalistic evidence from ADAS-supported production vehicles, this study contributes to understanding how external ADAS status visibility may shape following drivers’ longitudinal driving behaviors in real traffic.