Fig. 3: Performance evaluation of the dense learning approach on the nuPlan benchmark.
From: Breaking through safety performance stagnation in autonomous vehicles with dense learning
a Illustration of the four cities in the nuPlan benchmark where we evaluated the performance of our approach. b Architecture of the PDM-Hybrid with SafeDriver. SafeDriver uses an attention-based state dropout encoder and generates 8-second trajectories for vehicle control. c–f Two cases to demonstrate the effectiveness of our approach. The ego vehicle is depicted by a white rectangle. When controlled by SafeDriver, the ego vehicle is highlighted with a red circle. Other vehicles are shown using green rectangles, while pedestrians are represented by blue rectangles. The expert trajectory is marked by an orange curve. c When controlled by the base model, the ego vehicle proceeds straight ahead and collides with a vehicle turning right that does not yield. d In the same scenario, SafeDriver executes a hard brake to avoid the potential collision. e The base model directs the ego vehicle through a crosswalk, resulting in a collision with pedestrians. f In the same scenario, when encountering pedestrians, SafeDriver executes a proactive yielding, providing enough space to avoid the crash.
