Source-linked AI summary
KING: Generating Safety-Critical Driving Scenarios for Robust Imitation via Kinematics Gradients
Niklas Hanselmann, Katrin Renz, Kashyap Chitta, Apratim Bhattacharyya, Andreas Geiger
TL;DR
The paper addresses limited and naïve simulator traffic behavior that makes safety-critical scenarios difficult to generate and validate. KING optimizes background trajectories through a kinematic bicycle-model proxy, then uses expert solutions for imitation-learning fine-tuning. It achieves higher scenario-generation success than black-box optimization and supports improved collision robustness on generated and hand-crafted scenarios.
Problem
Current simulators have naïve background traffic and insufficient diversity, while real-world safety-critical data collection and validation are dangerous and not scalable.
Method
KING uses gradients through a kinematic bicycle model to perturb background traffic, and generated scenarios are used with expert behavior for imitation-learning fine-tuning.
Results
KING finds over 20% more safety-critical perturbations than BBO baselines and mixed-data fine-tuning improves Town10 driving score by over 3 points while reducing collision rate.
Takeaways & Limitations
Generated scenarios are challenging, diverse, solvable, and improve robustness on both KING scenarios and traditional hand-crafted scenarios.
Takeaways & Limitations
The generated scenarios target long-tail traffic and are not all likely to occur frequently in daily traffic.
Abstract
from arXiv · showhide
Simulators offer the possibility of safe, low-cost development of self-driving systems. However, current driving simulators exhibit naïve behavior models for background traffic. Hand-tuned scenarios are typically added during simulation to induce safety-critical situations. An alternative approach is to adversarially perturb the background traffic trajectories. In this paper, we study this approach to safety-critical driving scenario generation using the CARLA simulator. We use a kinematic bicycle model as a proxy to the simulator's true dynamics and observe that gradients through this proxy model are sufficient for optimizing the background traffic trajectories. Based on this finding, we propose KING, which generates safety-critical driving scenarios with a 20% higher success rate than black-box optimization. By solving the scenarios generated by KING using a privileged rule-based expert algorithm, we obtain training data for an imitation learning policy. After fine-tuning on this new data, we show that the policy becomes better at avoiding collisions. Importantly, our generated data leads to reduced collisions on both held-out scenarios generated via KING as well as traditional hand-crafted scenarios, demonstrating improved robustness.
1 Introduction
KING addresses limited and non-adaptive safety-critical scenario coverage by using gradients through a differentiable kinematics proxy to perturb background traffic, then supports robustness-oriented fine-tuning.
- 1 Introduction: Current simulators use naïve background behavior, while hand-crafted scenarios are non-adaptive and unlikely to cover the long tail of critical situations.These limitations motivate actively searching for agent-specific failure modes.
- 1 Introduction: KING uses gradients through a differentiable kinematics model to optimize non-critical initial scenarios into safety-critical perturbations.The method asks whether the full simulator must be differentiable and uses a kinematic bicycle model as a proxy for the simulator’s true dynamics.
- 1 Introduction: 20% more scenarios achieve safety-critical perturbations with KING than with black-box optimization.The comparison concerns the number of scenarios in which a valid safety-critical perturbation is found.
- 1 Introduction: Fine-tuning on expert behavior from safety-critical perturbations improves driving-agent robustness to collisions.Figure 1 presents scenario generation and subsequent fine-tuning as the two-stage workflow.
2 Related Work
Prior autonomous-driving research includes imitation-learning agents that either infer actions directly from sensors or use interpretable intermediate representations. Safety-critical scenario generation has largely used black-box optimization over parameterizations ranging from agent states to routes and policy components.
- End-to-End Driving: Imitation-learning driving agents either infer actions directly from raw sensor observations or operate on interpretable intermediate representations such as BEV semantic occupancy grids.The study considers AIM-BEV, which uses ground-truth BEV semantic occupancy, and TransFuser, which uses camera and LiDAR observations.
- End-to-End Driving: The paper evaluates two imitation-learning agents representing these paradigms: the AIM-BEV planner and the camera-and-LiDAR-based TransFuser.AIM-BEV acts on ground-truth perception, whereas TransFuser is an end-to-end agent.
- Generating Safety-Critical Scenarios: Earlier safety-critical scenario-generation methods use black-box optimization over variables including adversarial-agent states, route graphs, and driving-policy parameters.AdvSim instead parameterizes the search space as sequences of kinematic bicycle-model states for each adversarial agent.
3 Safety-Critical Scenario Generation for Robust Imitation
KING generates safety-critical perturbations of regular traffic by optimizing adversarial-agent trajectories with gradients through a differentiable kinematic model, then uses solvable scenarios to augment imitation-learning training.
- Driving Agents: The approach evaluates both a BEV planner and an end-to-end camera-and-LiDAR agent, using their policies within the scenario-generation and training pipeline.AIM-BEV uses a rasterized BEV representation, while TransFuser uses sensor observations and predicts future waypoints.
- Robust Training: The robust-training pipeline filters generated scenarios for safe expert alternatives, collects expert observation–waypoint pairs, and fine-tunes the policy on mixed critical and regular data.The generated scenarios augment the original training distribution with targeted safety-critical examples.
- Safety-Critical Perturbation: KING generates safety-critical perturbations by optimizing adversarial-agent actions while unrolling a differentiable kinematic bicycle model.The search space consists of action sequences for adversarial agents, and the model provides physically plausible motion while enabling backpropagation.
- Safety-Critical Perturbation: The simulation objective encourages ego–adversary collisions while penalizing adversary–adversary collisions and departures from drivable areas.These costs balance inducing failure in the driving policy with maintaining physically plausible scenarios.
- Safety-Critical Perturbation: Generated scenarios target the long tail of traffic and are not necessarily frequent, while the paper reports that they remain diverse, solvable, and useful for robust behavior learning.The realism of scenarios depends on the regularizing terms in the optimization objective.
- Kinematics Gradients: KING approximates true simulation gradients by retaining only the direct kinematics path and stopping gradients through the policy and renderer.This avoids requiring differentiable rendering or policies and reduces the computational overhead of backpropagating through those components.
4 Experiments
KING reliably generates diverse, solvable safety-critical scenarios and outperforms black-box optimization, while exposing failures in current imitation-learning agents. Training on KING-generated scenarios improves collision avoidance on both held-out generated cases and hand-crafted CARLA scenarios.
- 4.2 Comparison to BBO for Safety-Critical Scenario Generation: KING increases the number of scenarios with safety-critical perturbations by over 20% versus BBO baselines across all three traffic settings.It also outperforms baselines in convergence time as search dimensionality increases.
- 4.2 Comparison to BBO for Safety-Critical Scenario Generation: Using only direct gradients gives better results under a fixed computation budget than backpropagating through both direct and indirect simulator paths.The indirect path adds computational overhead despite reliably finding safety-critical perturbations.
- 4.3 Analysis of Safety-Critical Scenarios: KING generates diverse, solvable failure modes, including cut-ins, rear-ends, head-ons, merges, and intersection collisions, with only around 20% of scenarios deemed unsolvable.Failure cases are distributed across six clusters, while the rule-based expert solves challenging scenarios using privileged motion forecasting.
- 4.4 Evaluating Robustness after Fine-Tuning: Fine-tuning on combined regular and safety-critical data reduces collision rate to 28.57% on held-out KING scenarios and improves Town10 driving score by over 3 points.The combined dataset preserves a similar route-completion rate while reducing collisions on hand-crafted scenarios.
5 Conclusion
KING advances safety-critical traffic-scenario generation through efficient gradient guidance and improves collision avoidance when its scenarios augment imitation-learning data.
- KING uses a compute-efficient direct gradient path through a kinematic motion model to guide adversarial scenario generation.
- KING achieves higher success rates than existing black-box approaches while requiring less optimization time.
- Augmenting training data with KING-generated scenarios significantly improves collision avoidance for imitation-learning policies.