Technical Summary: ScenePilot

Problem Statement

Safety-critical scenarios are essential for evaluating autonomous driving (AV) systems but are exceedingly rare in naturalistic driving logs, necessitating simulation-based stress testing. Existing scenario generation methods typically treat surrounding agents as adversaries trained to induce collisions. However, these approaches suffer from two primary limitations:

1. Physical Invalidity: Methods that do not explicitly model vehicle-road physical limits often generate "visually extreme" but physically unsolvable crashes (e.g., requiring infinite deceleration). These failures test numerical robustness rather than revealing genuine deficiencies in the autonomy stack.
2. Over-Reliance on Single Constraints:
   • Physics-constrained methods restrict agents to physically realizable behaviors but often collapse onto extreme boundary-pushing maneuvers near physical limits, missing the broader regime of interactions that are physically solvable yet still cause AV failures.
   • Policy-constrained methods steer generation toward the capability boundary of a specific controller. While actionable for diagnosing that specific stack, they anchor generation to a controller-dependent boundary and fail to reveal how close the AV operates to the fundamental physical limits of the environment.

The paper argues that the most informative regime is the "boundary band": scenarios that are physically solvable in principle (admissible dynamics exist to avoid collision) but still cause the deployed autonomy stack to fail due to its specific limitations.

Methodology: ScenePilot

ScenePilot is a feasibility-guided, boundary-driven framework designed to target this boundary band. It formulates scenario generation as a constrained multi-objective Markov Decision Process (MDP) using Reinforcement Learning (RL).

1. Dual-Signal Formulation

The framework disentangles physical feasibility from AV policy capability using two per-step signals:

• Physical Feasibility Score ($\sigma$): Derived from a physics-limit variant of the Responsibility-Sensitive Safety (RSS) model. Unlike standard RSS which includes conservative reaction times, this variant calculates the theoretical braking-limit distance based on maximal feasible deceleration.
  • $\sigma \geq 0$: The frame is physically feasible (a collision is avoidable via admissible braking/steering).
  • $\sigma < 0$: The frame is physically infeasible (collision is unavoidable).
  • The score aggregates longitudinal and lateral margins, accounting for vehicle geometry and time-to-collision (TTC), to determine if orthogonal maneuvers can prevent a collision.
• AV-Risk Predictor ($\Phi$): An online-learned, lightweight critic network that estimates the instantaneous safety risk of the ego vehicle. It is trained via potential-based reward shaping to provide a dense, differentiable signal of collision probability, independent of a specific controller's policy.

2. Constrained Multi-Objective Optimization

The goal is to maximize AV risk ($J_\phi$) while maintaining physical feasibility ($J_\sigma \geq \epsilon$). The objective is formulated as:
$$\max_{\pi} J_\phi(\pi) \quad \text{s.t.} \quad J_\sigma(\pi) \geq \epsilon$$
where $\epsilon$ is a control variable that sweeps the feasibility threshold.

3. Key Mechanisms

• Step-Level Feasibility-Aware Shielding: To prevent the policy from drifting into physically infeasible regions or ignoring near-boundary violations, the authors introduce a shielding mask. If a feasibility violation is present or imminent ($h_t > 0$), the policy update prioritizes the feasibility advantage ($A_\sigma$); otherwise, it optimizes for risk ($A_\phi$). This ensures the agent stays near the boundary without collapsing into infeasible artifacts.
• Feasibility-Threshold Sweeping: The system employs a Gaussian-shaped schedule for $\epsilon$, starting near the physical limit and gradually increasing the margin. This allows the agent to explore a spectrum of the boundary band rather than collapsing onto a single narrow set of extreme maneuvers.

Key Contributions

1. Boundary-Band Formulation: A novel formulation that disentangles physical feasibility from AV policy capability, explicitly targeting the regime where interactions are physically solvable but policy-infeasible.
2. Constrained Adversarial Generator: A multi-objective RL framework coupling a physics-derived score ($\sigma$) and an online-learned risk predictor ($\Phi$), utilizing step-level shielding and threshold sweeping to concentrate samples in the near-boundary feasible band.
3. Empirical Validation: Demonstration that ScenePilot generates scenarios with higher collision rates while preserving physical validity, and that adversarial fine-tuning on these scenarios consistently improves downstream robustness.

Experimental Results

Experiments were conducted on the SafeBench benchmark within the CARLA simulator.

• Safety-Criticality: Compared to state-of-the-art baselines (including Learning-to-Collide, AdvSim, ChatScene, and SCENGE), ScenePilot achieved the highest mean Collision Rate (CR) of 0.893 and the lowest Overall Score (OS) of 0.476 across eight base traffic scenarios.
• Transferability: The generated scenarios remained effective across heterogeneous AV stacks, including CARLA Autopilot, AIM-BEV, TransFuser, BehaviorAgent, and various deep RL controllers (SAC, PPO, TD3). ScenePilot increased the average CR from 0.380 (baseline KING) to 0.490 on the Unprotected Left-turn scenario.
• Physical Validity: Analysis showed ScenePilot produced significantly fewer physically invalid frames compared to ChatScene, confirming that its failures are not dependent on physically impossible behaviors.
• Downstream Utility: Adversarial fine-tuning of an ego policy using ScenePilot-generated scenarios reduced the downstream collision rate from 0.854 (pre-trained) to 0.072, outperforming fine-tuning with scenarios from other generators.
• Coverage: The method covered a broader portion of the "physically feasible yet AV-infeasible" gap band (measured by Gap Coverage Score) compared to baselines, avoiding the collapse into overly mild or extreme regimes.

Significance and Claims

The paper claims that ScenePilot provides a more rigorous evaluation tool for autonomous driving by isolating competence gaps—failures that arise from limitations of the current autonomy stack rather than inevitable physical constraints. By focusing on the "boundary band," the method generates scenarios that are:

• Informative: They reveal genuine algorithmic weaknesses rather than numerical edge cases.
• Actionable: They provide targeted training data that improves the robustness of deployed systems.
• Generalizable: They remain challenging across diverse controller architectures.

The authors position this work as a step toward systematic safety assurance that moves beyond simple failure maximization to explore the nuanced space between physical possibility and algorithmic capability.