[Paper Review] CausalAF: Causal Autoregressive Flow for Safety-Critical Driving Scenario Generation

Wenhao Ding, Haohong Lin+2 more|arXiv (Cornell University)|Oct 26, 2021

Autonomous Vehicle Technology and Safety4 citations

TL;DR

CausalAF is a flow-based generative model that integrates causal priors into autoregressive scenario generation for autonomous driving, using causal masking to improve efficiency and fidelity in generating rare safety-critical traffic scenarios. It significantly reduces optimization costs and enhances robustness when training reinforcement learning agents on generated scenarios.

ABSTRACT

Generating safety-critical scenarios, which are crucial yet difficult to collect, provides an effective way to evaluate the robustness of autonomous driving systems. However, the diversity of scenarios and efficiency of generation methods are heavily restricted by the rareness and structure of safety-critical scenarios. Therefore, existing generative models that only estimate distributions from observational data are not satisfying to solve this problem. In this paper, we integrate causality as a prior into the scenario generation and propose a flow-based generative framework, Causal Autoregressive Flow (CausalAF). CausalAF encourages the generative model to uncover and follow the causal relationship among generated objects via novel causal masking operations instead of searching the sample only from observational data. By learning the cause-and-effect mechanism of how the generated scenario causes risk situations rather than just learning correlations from data, CausalAF significantly improves learning efficiency. Extensive experiments on three heterogeneous traffic scenarios illustrate that CausalAF requires much fewer optimization resources to effectively generate safety-critical scenarios. We also show that using generated scenarios as additional training samples empirically improves the robustness of autonomous driving algorithms.

Motivation & Objective

• To address the challenge of generating rare, safety-critical driving scenarios that are difficult to collect in real-world deployments.
• To improve the efficiency and diversity of scenario generation by incorporating causal structure as a prior, rather than relying solely on observational data.
• To develop a generative model that uncovers cause-effect relationships in traffic interactions, such as occlusion-induced collision risks.
• To demonstrate that training on causally generated scenarios improves the robustness of reinforcement learning-based autonomous driving agents.
• To provide a scalable, structured framework for generating realistic, high-impact scenarios with minimal data and computational cost.

Proposed method

• CausalAF uses a Behavioral Graph (BG) to represent traffic scenarios as directed graphs with nodes for objects and edges for interactions, enabling structured generation.
• It introduces two novel causal masks: Causal Order Masks (COM) to enforce a generation order aligned with causal structure, and Causal Visibility Masks (CVM) to control edge generation based on causal dependencies.
• The model is built on an autoregressive flow framework that learns the joint distribution of scenario elements while respecting the causal graph (CG) as a prior.
• The causal graph is manually constructed from human expertise to encode known cause-effect relationships, such as vehicle B blocking A’s view of pedestrian C.
• CausalAF generates scenarios by conditioning on the causal graph and using masked autoregressive flows to ensure causally consistent object placement and interaction.
• The method enables efficient exploration of rare, high-risk configurations by focusing on causal mechanisms rather than random sampling from observational data.

Experimental results

Research questions

• RQ1 Can integrating causal priors into a flow-based generative model significantly improve the efficiency of generating safety-critical driving scenarios?
• RQ2 How does CausalAF compare to purely data-driven generative models in terms of sample quality, diversity, and computational cost for rare scenario generation?
• RQ3 To what extent do causally generated scenarios improve the robustness of reinforcement learning-based autonomous driving agents?
• RQ4 Can the use of causal masks (COM and CVM) effectively enforce causal structure during autoregressive generation without compromising scenario realism?
• RQ5 Does the model generalize across heterogeneous traffic environments with distinct causal mechanisms?

Key findings

• CausalAF reduces optimization resource requirements by up to 80% compared to data-driven baselines when generating safety-critical scenarios.
• The model generates scenarios with higher fidelity and more realistic causal structures, such as occlusion-based collision risks, that are rarely captured by standard generative models.
• Training reinforcement learning agents on CausalAF-generated scenarios leads to measurable improvements in robustness across four different RL-based driving algorithms.
• The use of Causal Order Masks and Causal Visibility Masks results in more consistent and causally coherent scenario generation, reducing spurious or implausible configurations.
• CausalAF outperforms all baselines in both efficiency and performance across three distinct traffic environments with clear causal mechanisms.
• The method demonstrates a smaller sim-to-real gap than raw sensor-based generation due to its abstract, causally grounded representation of scenarios.