A learning narrow-passage assist: An agentic design case study
An interaction-design study of an in-car cluster HMI for a Level-4 vehicle that learns from repeated driving situations. Because the car drives itself, the interface’s job shifts from instructing a driver to earning a passenger’s trust through transparent, adaptive communication. The worked example is a narrow construction-zone passage where the car must pass a vehicle in the adjacent lane with little lateral clearance and the HMI behaves differently the first time it meets the situation, after it has handled it many times, and when something unexpected breaks the pattern. A functional prototype was built in Unity on the Unity Automotive HMI Template.
Rather than a static dashboard that mirrors vehicle signals, the HMI is modelled as an agent. It perceives the environment and the passenger, reasons about risk and intent, acts by communicating, and learns from the outcome. The defining behavior is learning from repeated patterns: the system recognizes recurring contexts a geofenced roadworks zone met on a daily commute and adapts both how much it communicates and how cautiously it drives.
Agency here is not full self-direction; it is a bounded loop with goals (safety, trust, comfort, energy) and a memory that lets experience change behavior over time.
The HMI runs a continuous perceive → reason → act → learn loop around a central experience memory. Perception fuses environment and passenger data; reasoning classifies risk and selects intent; action drives the visualization and narration; learning records the outcome. The experience memory is the across-time element the learn stage writes to it and the reason stage recalls from it and it is what turns a sequence of identical loops into adaptive behavior.
The prototype runs in Unity 6 (6000.4.10f1) using the Universal Render Pipeline and the Unity Automotive HMI Template. All behavior was added in custom C# alongside the template, without modifying template scripts.
NarrowPassageAssist > Orchestrates the maneuver; drives road/speed; choreographs the pass
SafeGapCorridor > Measures lateral clearance; colors and sizes the corridor marker
ChaseCamera > Elevated 3/4 follow camera for the maneuver view
LearningState > Switches naïve / calibrated / anomaly; drives prominence, speed, HUD
SlopeChargeInfluencer > Injects regen energy to show downhill charging
Concept, design and prototype: Shiva Mohseni. Base: Unity Automotive HMI Template. References: SAE J3016 (levels of automation); Lee & See (trust in automation); agent-transparency (SAT) model.
