Safety

Pioneering Safety 2.0: A New Paradigm in AV Safety

Rethinking AV safety: Embracing a new approach

Scaling AVs to widespread use has faced significant challenges, particularly when confronted with the ‘long tail’ of rare and unpredictable driving scenarios. Traditional methods attempt to tackle these challenges through extensive scenario modeling and rigid safety frameworks. However, this approach falls short, as it fails to replicate the nuanced understanding humans acquire through years of driving—a deep, intuitive grasp of navigating complex situations that goes beyond any set of fixed rules.

Introducing Safety 2.0

A revolutionary approach that addresses safety through the lens of deep world understanding. Unlike traditional methods, Safety 2.0 acknowledges that true safety comes from an AI that interprets the driving environment naturally, like a human driver. It establishes an AI system that inherently comprehends the intricacies of the world and driving behaviors, which it uses to navigate safely.

Learning to predict and adapt

Utilizing Generative AI technologies, Wayve has developed a high-fidelity, predictive world model that understands the implications of the vehicle’s actions and ensures safe responses to other road users. The model’s understanding can be used to prevent unsafe actions.

Embedding safety into the AI’s core

Wayve’s end-to-end (e2e) model architecture and active learning processes are optimized to produce effective, fluent, and safe driving behaviors from unlabeled driving videos, showcasing strong generalization across various geographies, vehicles, and sensors. This can be enhanced with multimodal data sources that provide additional information.

Building an emergency response safety reflex

Emergency response capabilities are incorporated into the AI model by boosting its exposure to critical scenarios. This approach accelerates the learning curve in the sub-domain of emergency response, allowing us to cultivate an innate safety reflex that improves driving safety.

Safety through predictable, cost-efficient training

Wayve’s Safety 2.0 process emphasizes the development of ‘introspectable’ models that can be systematically examined and reviewed. This enables effective monitoring and guidance of the AI model’s unsupervised learning outcomes, ultimately transforming large datasets into actionable insights that lead to predictable and regression-resistant performance enhancements.

Scenario intelligence

Creating new methods for dataset introspection and control. These tools harness the emergent concepts generated by the AI and transform them into a framework for deeper analysis.

Data quality and control

Creating comprehensive datasets for diverse driving scenarios and filtering out low-quality and risky scenarios.

Model introspection

Pioneering methods for model introspection using natural language and other information to evaluate the AI’s decision-making process.

Model deployment dashboard

Providing OEMs with control and visibility over their data assets and learning process to ensure that Wayve’s AI models achieve the highest safety standards.

Aligning with established Safety Standards

Safety 2.0 strategy aligns with critical safety principles, including redundancy, fail-operational behaviors, and adherence to automotive safety standards such as ISO 26262 and SOTIF. It also ensures scalability and commercial viability by leveraging economical data, scalable cloud infrastructure, and partner fleets.

Leveraging synthetic data

Wayve’s GAIA-1 generates realistic driving scene videos with high controllability. This plentiful source of synthetic data efficiently fills data gaps around edge cases.

Expanding through large partner fleets

Quickly acquiring diverse driving data through collaborations with partner fleets.

Utilizing scalable cloud infrastructure

Collaborating with Microsoft for supercomputing infrastructure to support AI training and data workloads on a global scale.

Learn more about Wayve's Safety Initiatives

Back to top