A new AI framework can predict whether a specific driver will brake or run a yellow light — and when they will decide.
Researchers introduced VISTA-DZ, a system that converts a driver's historical approach trajectories into visual representations, then feeds those visuals to a vision-language model to generate a plain-language behavioral profile for each driver. That profile gets encoded as a semantic embedding and used to condition a prediction network built on a bidirectional GRU encoder, multi-head cross-attention, and a feature modulation layer. The result: the model knows not just what the road looks like, but who is driving. Tested on two datasets — one simulated, one newly collected in the field — VISTA-DZ hit 93.26% accuracy in simulation and 90.22% mean accuracy across 20 drivers it had never seen before.
The dilemma zone problem is deceptively hard. Two drivers approaching the same intersection at the same speed can make opposite calls based on risk tolerance, braking habit, and personal thresholds — factors that scalar summaries like average deceleration have always struggled to capture. Encoding behavior as language and then reasoning over it is a meaningful step beyond those handcrafted descriptors, and the zero-shot simulation-to-real transfer results suggest the approach generalizes without requiring mountains of real-world data.
Adaptive signal controllers and driver-assistance systems have long needed personalization that scales. Whether a vision-language detour is the most efficient path to get there — or just the trendiest one — is a question the deployment numbers will eventually answer.