TLDR: SAGE (Steerable Adversarial scenario GEnerator) is a new framework for creating challenging yet realistic scenarios to test autonomous driving systems. Unlike previous methods that offer a fixed balance between adversariality (how challenging a scenario is) and realism (how plausible it is), SAGE allows users to adjust this balance in real-time without retraining. It achieves this by training two ‘expert’ models (one for high adversariality, one for high realism) and then smoothly blending their internal settings at the moment of scenario generation. This approach, supported by theoretical principles like Linear Mode Connectivity, leads to more effective and robust training of autonomous driving policies, as demonstrated by extensive experiments.
Assessing the safety of autonomous driving (AD) systems is a critical step before they can be widely deployed. One of the most effective and cost-efficient ways to do this is through adversarial scenario generation. This involves creating challenging situations in simulations that push AD systems to their limits, helping identify vulnerabilities and improve their robustness.
However, existing methods for generating these scenarios often face a significant challenge: they are typically designed with a fixed balance between how ‘adversarial’ (challenging) a scenario is and how ‘realistic’ (plausible in the real world) it appears. This means that if you need a very aggressive scenario for stress testing, or a more realistic but still challenging one for training, you often have to retrain the entire model, which is both time-consuming and inefficient. These models lack the flexibility to adapt to diverse testing and training needs on the fly.
Introducing SAGE: Steerable Adversarial Scenario Generator
A new framework called SAGE (Steerable Adversarial scenario GEnerator) redefines this problem as a multi-objective preference alignment task. SAGE offers a groundbreaking solution by allowing fine-grained control over the trade-off between adversariality and realism at the time of testing, without any need for retraining. This means users can dynamically adjust how aggressive or realistic a generated scenario should be, tailoring it to specific requirements.
How SAGE Works: Hierarchical Group-based Preference Optimization (HGPO)
At the heart of SAGE is a novel data-efficient offline alignment method called Hierarchical Group-based Preference Optimization (HGPO). This method learns to balance competing objectives by cleverly separating ‘hard’ feasibility constraints (like ensuring a vehicle stays on the road or doesn’t collide with static objects) from ‘soft’ preferences (the desired balance of adversariality and realism). Unlike traditional methods that might penalize map violations as just another ‘bad’ outcome, HGPO treats them as absolute preconditions, ensuring that all generated scenarios are fundamentally valid.
Instead of training a single, fixed model, SAGE fine-tunes two specialized ‘expert’ models. One expert is optimized for high adversariality, and the other for high realism. At inference time – when a scenario needs to be generated – SAGE constructs a continuous spectrum of policies by simply interpolating the weights of these two experts. This linear interpolation allows for a smooth transition between highly realistic and highly adversarial behaviors, providing unparalleled control.
The theoretical justification for this approach comes from the concept of Linear Mode Connectivity (LMC). LMC suggests that when models are fine-tuned from the same starting point on related tasks, their optimal solutions often lie in a ‘flat’ region of the parameter space, allowing for effective linear interpolation between them without significant loss in performance.
Benefits and Applications
Extensive experiments demonstrate that SAGE excels in several ways:
- Superior Scenario Quality: SAGE generates scenarios that achieve a better balance of adversariality and realism compared to state-of-the-art methods. It can create highly challenging scenarios while maintaining significantly lower penalties for unrealistic behaviors and map violations.
- Test-Time Steerability: The framework allows users to smoothly adjust the adversarial weight, transitioning generated trajectories from compliant lane-following to aggressive cut-ins or sudden brakes. This dynamic control is crucial for targeted stress testing and data augmentation.
- Effective Closed-Loop Training: When integrated into a closed-loop adversarial training pipeline, SAGE helps improve driving policies more effectively. It uses a ‘dual-axis curriculum’ that progressively increases the intensity and frequency of adversarial encounters as the AD agent learns, preventing ‘catastrophic forgetting’ where the agent becomes good at handling rare adversarial cases but forgets how to drive normally.
The research paper, available at https://arxiv.org/pdf/2509.20102, provides detailed theoretical and empirical evidence supporting SAGE’s effectiveness. Ablation studies further confirm the importance of HGPO and the separation of hard constraints from soft preferences.
Also Read:
- Real-World Cooperative Perception for Safer Lane Changes: A Hardware Deployment Study
- Improving Safety and Interpretability in Autonomous Driving Through Entropy Regularization
Looking Ahead
SAGE establishes a new foundation for steerable scenario generation. Future work could explore incorporating a richer set of objectives (like scenario novelty or complexity), more advanced model merging techniques, and automated curricula that dynamically adapt based on the agent’s learning progress, leading to even more intelligent adversarial training for autonomous vehicles.
