Autonomous Driving Safety Performance on LLM-Attacker Scenarios
20Failure RateLLM-Attacker
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| LLM-AttackerScenario Family=CUT_IN_LEFT, Residual Failure Mode=The ego vehicle follows the leading vehicle too closely and fails to recognize the blocked path ahead. When the front vehicle turns abruptly, the ego continues straight and collides.2026.06 | 20 | 80 | |
| LLM-AttackerScenario Family=CUT_IN_RIGHT, Residual Failure Mode=The ego reacts too late to suddenly spawned vehicles and cannot brake in time.2026.06 | 20 | 80 | |
| LLM-AttackerScenario Family=WRONG_WAY_DRIVER, Residual Failure Mode=The policy responds only with braking and fails to execute evasive steering, resulting in head-on collision.2026.06 | 30 | 70 | |
| LLM-AttackerScenario Family=SUDDEN_BRAKE, Residual Failure Mode=The ego cannot decelerate quickly enough when a vehicle emerges abruptly ahead.2026.06 | 40 | 60 | |
| LLM-AttackerScenario Family=CYCLIST_CROSS, Residual Failure Mode=The cyclist intersects the middle of the ego vehicle body, and braking alone is insufficient for avoidance.2026.06 | 70 | 30 |