Breaking through safety performance stagnation in autonomous vehicles with dense learning.

Autonomous vehicles remain commercially limited largely due to safety performance stagnation. Existing deep learning, heavily reliant on failure data from rare safety-critical events, suffers from the seesaw effect-improvement in some scenarios causes regression in others. We introduce an innovative dense learning approach that prioritizes both informative failures and successes, informed by theoretical findings. Data is sampled proportionally to its contribution to the policy gradient and exposure frequency, excluding non-informative samples. This densifies the training dataset's information, significantly reducing learning variance without bias, enabling tasks intractable for existing methods. To validate this, we trained a safety-critical driving agent for a highly automated vehicle using mixed reality on an urban test track. Results demonstrate that our approach breaks the performance stagnation, enhancing the model's overall safety performance by one to two orders of magnitude. This marks a significant stride towards achieving human-level safety and widespread adoption for autonomous vehicles.