← Back ICRA 2026

SEAL: Towards Safe Autonomous Driving Via Skill-Enabled Adversary Learning for Closed-Loop Scenario Generation

Benjamin Stoler, Ingrid Navarro, Jonathan Francis, Jean Oh

PDF

AI summary

Key figure (auto-extracted from paper)

SEAL improves autonomous driving policy robustness by generating realistic, reactive adversarial scenarios that boost ego task success by over 20% compared to state-of-the-art baselines.

Autonomous Driving Scenario Generation Adversarial Training Closed-Loop Simulation Safety-Critical Testing Skill-Based Policies

Problem

Existing safety-critical scenario generation methods rely on simplistic objectives, producing overly aggressive or non-reactive adversaries that lack realism and fail to provide effective training stimuli for closed-loop ego policies.

Approach

SEAL integrates a learned objective function to anticipate ego responses with a hierarchical, skill-enabled adversarial policy that selects human-like maneuvers to generate realistic, reactive, and safety-critical scenarios.

Key results

Proposes a learned objective function predicting collision closeness and ego behavior deviation
Introduces a reactive, skill-based adversarial policy with adversarial/benign priors for human-like realism
Achieves over 20% higher ego task success rate across generated and real-world scenarios compared to SOTA baselines
Establishes an out-of-distribution evaluation framework using real-world safety-relevant scenarios

Why it matters

It enables more realistic and rigorous stress-testing for autonomous driving systems, accelerating the development of safer self-driving policies.

Abstract

Verification and validation of autonomous driving (AD) systems and components is of increasing importance, as such technology increases in real-world prevalence. Safety-critical scenario generation is a key approach to robustify AD poli- cies through closed-loop training. However, existing approaches for scenario generation rely on simplistic objectives, resulting in overly-aggressive or non-reactive adversarial behaviors. To generate diverse adversarial yet realistic scenarios, we propose SEAL, a scenario perturbation approach that leverages learned objective functions and adversarial, human-like skills. SEAL- perturbed scenarios are more realistic than SOTA baselines, leading to improved ego task success across real-world, in- distribution, and out-of-distribution scenarios, of more than 20%. To facilitate future research, we release our code and tools: https://navars.xyz/seal/

Index terms

Intelligent Transportation Systems Autonomous Vehicle Navigation Performance Evaluation and Benchmarking