Learned Safety: Building Trustworthy Embodied AI

Since 2017, Wayve has been developing end-to-end embodied AI that offers a fundamentally different path to solve autonomous driving’s long-tail problem. Rather than attempting to anticipate every possible situation through maps, pre-defined rules, and human-engineered representations, we believe that intelligence must learn to drive competently and generalize from experience.

Nearly a decade later, our work has advanced from that thesis to real-world proof as we bring the Wayve AI Driver to production around the world with our partners. Now there’s a more concrete question:

How do we build, observe, and validate safety within an end-to-end AI Driver at global scale?

infographic showing approach to learned safety

This is a fundamental challenge for embodied AI. Unlike AI systems that operate in a digital world, embodied intelligence must interact with a physical environment that is complex, dynamic and unpredictable. It must learn not only to understand its surroundings, but to make decisions and take actions that account for real-world risks. Autonomous driving is one of the most demanding applications of this challenge, where the consequences of an incorrect decision can be immediate and significant.

To date, the AV industry has often favored benchmarks that are easier to inspect, count, or explain as stand-ins for safety evidence: stacks divided into separate components from perception to planning, accumulated mileage and disengagement rates, or human-readable explanations of a system’s actions. Each can contribute to a safety case. None establishes safety on its own.

Learned Safety

Building Safety Into End-To-End AI

At Wayve, we believe safety must be part of how the AI Driver learns to understand and act in the physical world, rather than a checker or constraint imposed after the action’s been decided. Just like driving, embodied AI safety improves faster with learned representations than with anything hand-engineered.

But learning safe behavior is only part of the challenge. We must also make safety-relevant judgments observable, evaluate the complete driving system, and establish credible evidence that its residual risk is acceptable.

In this three-part blog series, we explore the principles and technologies behind Wayve’s approach to Learned Safety, the notion that, like driving, safety and validation are best solved with end-to-end deep learning.

Learned Safety

What this means for Embodied AI

The promise of embodied AI is intelligence that can learn to operate in the physical world, adapting to environments and situations it has never encountered before. Realizing its potential requires keeping safety central to how these systems learn and act, alongside verification and validation methods that scale with their capabilities.

Discover what Wayve is learning as we work towards our vision: an AI Driver that can safely drive any vehicle, anywhere.

Aligning with established Safety Standards

Wayve’s Learned Safety approach also aligns with critical safety principles, including redundancy, fail-operational behaviors, and adherence to automotive safety standards such as ISO 26262 and SOTIF.