#USA #New York #AI Safety #Nexar #BADAS 2.0

Nexar Unveils BADAS 2.0

On April 16, 2026, Nexar, a leading real-world intelligence platform, proudly announced the launch of BADAS 2.0, an advanced incident prediction model that sets a new benchmark in road safety technology. Developed using an extensive dataset that includes over two million real-world collision-risk events collected from 200 million miles of driving, BADAS 2.0 promises to revolutionize the way we predict and respond to road incidents.

A Leap Forward in Predictions

The latest iteration of the BADAS model builds upon its predecessor, BADAS 1.0, using the same V-JEPA2 architecture. However, BADAS 2.0 represents a significant enhancement, being trained on five times the amount of data. This allows the new model to outperform not only its forerunner but also a large, 2-billion-parameter foundational model, achieving 99.4% average precision across all benchmarks.

Zach Greenberger, CEO of Nexar, stated, "When you train on two million real collision-risk events from 200 million miles of road, the model doesn’t just score better; it reasons better. That’s a different thing entirely, and it’s only possible with data at this scale."

Three Models for Diverse Environments

BADAS 2.0 is equipped with a family of three distinct models, each significantly optimized for various deployment scenarios:
1. BADAS 2.0 (300M parameters): This large base model excels in performance on all major benchmarks, showcasing the best mean time-to-action and recall for early warnings, particularly adept in handling rare long-tail scenarios.
2. BADAS 2.0 Flash (86M parameters): Designed for end-user devices, this optimization focuses on preventing false alarms while surpassing BADAS 1.0 on all performance metrics.
3. BADAS 2.0 Flash Lite (22M parameters): An ultra-lightweight model tailored for IoT and edge deployments, Flash Lite can perform at speeds 12 times faster than the full model while maintaining nearly the same precision levels.

All three versions are designed to operate within a 66-millisecond real-time response time, making them practical for on-the-road applicability.

Enhanced Explainability and Reasoning

With the aim of turning alerts into actionable responses, BADAS 2.0 introduces critical capabilities:

• - Explainability: The model employs attention heatmaps that reveal what it perceives during risk situations, ensuring that decisions made for incident predictions are transparent and auditable.
• - Reasoning: Unlike traditional models, BADAS 2.0 doesn't stop at incident prediction. It comprehensively articulates the expected actions and the rationale behind them.
• - Generalization: This new model exhibits consistent performance in predicting incidents across diverse scenarios, including unique out-of-distribution contexts, thereby significantly improving safety outcomes on the road.

Harnessing Real-World Data

BADAS 2.0 was trained on a vast repository of classified naturalistic safety-critical events comprising 60 million edge-case videos derived from total travel data of 10 billion miles. This robust training set underscores BADAS 2.0's ability to handle challenging scenarios that have historically been problematic for other models, such as detecting animals on roads, navigating through fog, and operating in slick snowy conditions.

The original BADAS model, launched in late 2025, utilized a much smaller dataset of 400,000 real-life dashcam clips. The expansion to 2 million collision-risk events in BADAS 2.0 allows for a level of training that is unprecedented in the industry, powered by a network comprising 350,000 cameras that capture 100 million miles of driving each month.

In conclusion, BADAS 2.0 is not just an upgrade but a testament to Nexar's commitment to leveraging real-world data for enhanced road safety. With its advanced capabilities, BADAS 2.0 is poised to become an indispensable asset in the realm of incident prediction and road safety technology, now available to users worldwide.