How Well Do Vision-Language Models Understand Sequential Driving Scenes? A Sensitivity Study

VENUSS addresses a critical gap in the 'VLM for Robotics' space: the fact that most current models are optimized for static images rather than temporal sequences. However, as an academic project with 0 stars and 3 forks, it currently lacks the community momentum or proprietary data to form a moat. The defensibility is low (2) because it is primarily a research artifact for a paper; its value lies in its methodology rather than a sticky product or network effect. Frontier labs like Waymo (Alphabet), Tesla, and NVIDIA are already developing sophisticated internal temporal-VLM benchmarks that likely exceed the depth of this public framework. The project's strength is its focus on 'sensitivity analysis'—understanding how minor changes in input affect model output—which is a niche area labs sometimes overlook in favor of raw performance. Expect this to be superseded within 1-2 years as end-to-end driving models (like Wayve's or Tesla's FSD v12) integrate native temporal-linguistic reasoning directly into their training loops.