Learned Safety: Building Trustworthy Embodied AI by Wayve

Wayve has developed an end-to-end embodied AI system called the Wayve AI Driver, which focuses on learning to drive safely and generalize from experience rather than relying on pre-defined rules. The company believes safety should be integrated into the decision-making process rather than being checked afterward. The Wayve Learned Safety Envelope (LSE) is a proprietary system that incorporates safety judgments into decision-making and provides a heatmap to show where the vehicle is and isn’t safe to drive. This approach helps answer three key questions: how to act, what is safe, and whether the AI can continue with confidence.

The company’s Validation 2.0 approach uses AI scaling properties to improve validation with more data, vehicles, and driving domain diversity. It combines closed-loop simulation with GAIA, learned evaluation through a reward model, GRADER, and statistical evidence transfer to turn observable behavior into a rigorous safety argument. This method addresses the challenge of validating safety in complex, dynamic environments.

Learned Safety is part of how the AI Driver learns to understand and act in the physical world. It builds on the idea that safety and validation are best solved with end-to-end deep learning. The approach aligns with safety standards like ISO 26262 and SOTIF, and leverages synthetic data through GAIA to evaluate safety in critical situations. The goal is to create an AI Driver that can safely drive any vehicle, anywhere, by keeping safety central to how the system learns and acts.

AI-generated summary of the original article — see the source for the full story.