Kinematics-Based Evaluation of Vision-Language Models for Cyclist Right-Turn-on-Red Advisories

The French M12 right-turn-on-red rule improves cycling efficiency but increases interaction risk at signalized intersections. We propose a kinematics-based digital-twin benchmark to evaluate the safety and regulatory compliance of Vision-Language Models (VLMs) as cyclist assistants. A SUMO–Unity co-simulation generates synchronized dual-view renderings and kinematic data, while a conservative Conflict Detection Model (CDM) provides ground-truth advisories via geometric screening and strict gap acceptance. We benchmark four VLMs (Gemini-3.0 Pro/Flash, Qwen3-VL Plus/Flash) on 300 scenarios under systematic input ablations. Results show significant safety disparities: Qwen3-VL exhibits stronger regulatory grounding and becomes fail-safe when the M12 exemption text is removed, whereas Gemini models display riskier visual biases. Bird’s-eye context is essential; first-person-only inputs yield either unsafe over-optimism or excessive stopping. Although all models achieve zero explicit regulatory violation, critical failures remain frequent due to visual misinterpretations of Traffic lights and signage. Our findings suggest VLM-based cyclist assistants require deterministic perception and physics-based guardrails before safety-critical deployment.