Risk-Aware World Model Predictive Control for Generalizable End-to-End Autonomous Driving
Proposes RaWMPC, a risk-aware predictive control framework that learns safety without expert supervision, improving autonomous driving generalization.
Key Findings
Methodology
RaWMPC constructs a transformer-based world model to predict future states and traffic events. It employs a risk-aware interaction strategy, systematically exposing the model to hazardous behaviors through online simulation. The framework integrates a contrastive learning-based self-evaluation distillation, transferring risk avoidance capabilities into a lightweight action proposal network. Training combines offline trajectory warm-up and online environment interaction, optimizing for robustness and safety. The core components include state prediction, risk assessment, and decision optimization, validated on CARLA and NAVSIM, outperforming state-of-the-art methods especially in out-of-distribution scenarios.
Key Results
- On NAVSIM and Bench2Drive, RaWMPC achieved over 15% reduction in collision rates in unseen scenarios, with an average collision rate dropping to 3.2%. It demonstrated superior robustness in complex traffic situations, significantly outperforming imitation learning and model-based RL baselines.
- Ablation studies confirmed that risk-aware interaction training and self-evaluation distillation contributed substantially to performance gains. The approach improved risk prediction accuracy and generated safer candidate actions, enhancing decision transparency and reliability.
- The method enables autonomous systems to learn risk avoidance purely through environment interaction, reducing dependence on costly expert data. It shows promise for real-world deployment, offering scalable, interpretable, and safer autonomous driving solutions.
Significance
This work addresses the core challenge of generalization in autonomous driving by shifting from imitation to proactive risk-aware control. It enables autonomous vehicles to handle rare and hazardous scenarios without relying on extensive expert data, significantly advancing safety and reliability. The risk-aware interaction and distillation mechanisms provide a new paradigm for autonomous systems, emphasizing active risk prediction and mitigation. This approach enhances decision interpretability and robustness, bridging the gap between simulation and real-world deployment, and paving the way for safer autonomous vehicles in complex environments.
Technical Contribution
RaWMPC introduces a novel risk-aware predictive control framework that integrates a transformer-based world model with explicit risk evaluation. It innovatively employs risk-aware interaction to systematically expose the model to hazardous behaviors, improving its prediction of dangerous outcomes. The self-evaluation distillation transfers complex risk-avoidance skills into an efficient action proposal network, enabling real-time deployment. Unlike prior methods relying heavily on expert demonstrations, this approach emphasizes environment-driven risk learning, providing theoretical guarantees of safety improvement and enhanced generalization in unseen scenarios.
Novelty
This is the first work to integrate risk-aware interaction strategies with world model-based predictive control for autonomous driving, enabling active learning of hazardous scenarios without expert labels. The combination of risk exposure, explicit risk evaluation, and distillation into a lightweight policy network offers a new paradigm that surpasses traditional imitation and reinforcement learning approaches in safety and generalization. The approach fundamentally shifts the focus from mimicking expert behavior to proactive risk management, representing a significant innovation in autonomous decision-making.
Limitations
- The framework's performance depends on the quality of simulated hazardous interactions; in real-world deployment, unmodeled environmental complexities may limit effectiveness.
- Training involves high computational costs due to extensive environment interaction and model optimization, posing challenges for real-time applications.
- Sensor noise and environmental uncertainties in real vehicles could impair the accuracy of risk prediction, requiring further robustness enhancements.
Future Work
Future research will focus on integrating multi-modal sensor data, such as radar and camera fusion, to improve risk perception robustness. Developing online adaptation mechanisms for real-world environments and extending the framework to multi-agent scenarios will be key directions. Additionally, efforts will be made to reduce computational overhead and enhance real-time safety guarantees, facilitating deployment in commercial autonomous vehicles.
AI Executive Summary
Autonomous driving has seen rapid advancements through imitation learning and large-scale datasets. However, these methods often struggle to generalize safely in rare or unseen scenarios, limiting their deployment in real-world environments. Traditional approaches rely heavily on mimicking expert demonstrations, which do not cover all possible traffic situations, especially hazardous ones. This gap raises the need for systems capable of proactive risk assessment and avoidance.
In response, the authors introduce RaWMPC, a risk-aware world model predictive control framework that eliminates the dependence on expert supervision. The core idea is to leverage a transformer-based world model to predict the consequences of multiple candidate actions and evaluate their associated risks explicitly. By systematically exposing the model to hazardous behaviors through online simulation, RaWMPC enables the system to learn the consequences of risky maneuvers, thus improving safety and robustness.
A key innovation is the self-evaluation distillation process, which transfers the risk-avoidance skills learned during interaction into a lightweight action proposal network. This allows for efficient, real-time decision-making without sacrificing safety or interpretability. Extensive experiments on simulated datasets such as CARLA and NAVSIM demonstrate that RaWMPC outperforms existing state-of-the-art methods, especially in out-of-distribution scenarios, reducing collision rates by over 15% and enhancing decision transparency.
This approach marks a significant step toward safer autonomous vehicles capable of handling complex, unpredictable environments. By focusing on active risk prediction and proactive avoidance, RaWMPC addresses fundamental limitations of prior methods, paving the way for more reliable and generalizable autonomous driving systems. Future work will aim to incorporate multi-modal data, improve real-world robustness, and extend to multi-agent settings, ultimately bringing autonomous vehicles closer to safe, widespread deployment.
Deep Analysis
Background
Autonomous driving技术经历了从基于规则到深度学习的演变。早期方法如Thrun等的强化学习和规则系统，强调感知与决策的分离。近年来，模仿学习（IL）和模型基强化学习（MBRL）成为主流，利用大规模数据集（如Waymo Open Dataset、nuScenes）推动性能提升。然而，这些方法在长尾风险场景中表现不足，难以应对未见或极端环境。现有研究多依赖专家示范，缺乏主动风险建模和规避机制，导致在未知场景中的安全性不足。近年来，结合世界模型和预测控制的研究逐渐兴起，但大多仍受限于对专家数据的依赖和泛化能力。
Core Problem
核心问题在于，现有端到端自主驾驶模型在面对未见或极端场景时，容易出现不安全行为。模仿学习依赖专家示范，难以覆盖所有潜在风险，导致模型在长尾场景中表现不佳。强化学习虽能自主探索，但缺乏对高风险事件的明确建模，难以保证安全。如何在不依赖专家示范的情况下，主动学习危险行为的后果，并实现风险规避，成为亟待解决的难题。解决这一问题对于提升自主系统的安全性、可靠性和泛化能力具有重要意义。
Innovation
本研究的创新点包括：1）引入风险感知交互策略，系统性暴露模型于危险行为，提升其风险预测能力；2）利用世界模型进行多候选行为的预测与风险评估，主动规避高风险行为；3）设计自我评估蒸馏机制，将风险规避能力迁移到轻量级动作网络，实现高效推理。不同于传统模仿学习仅复制专家行为，RaWMPC强调主动风险规避，增强模型的鲁棒性和泛化能力。这一机制在自主驾驶领域尚属首次，突破了现有方法对专家示范的依赖，提供了全新的自主学习路径，具有重要的理论创新和实践价值。
Methodology
- �� 构建基于Transformer的世界模型，预测未来状态和交通事件。
- �� 利用多模态输入（视觉、ego状态）编码，生成状态和行为的潜在表示。
- �� 通过风险感知交互策略，系统性暴露模型于危险行为，训练其预测危险后果。
- �� 在训练中结合离线轨迹预热和在线模拟交互，优化模型鲁棒性。
- �� 设计多候选行为预测与风险评估机制，选择低风险行为。
- �� 引入自我评估蒸馏，将风险规避能力迁移到动作生成网络，提升推理效率。
Experiments
采用CARLA和NAVSIM模拟环境，评估模型在多场景下的表现。比较基线包括模仿学习和模型基强化学习，指标涵盖碰撞率、决策透明度和泛化能力。超参数如预测步长H、风险阈值和温度参数调优，进行消融实验验证各模块贡献。模型在未见场景中实现了15%以上的安全性提升，碰撞率降至3.2%，验证了风险感知策略的有效性。
Results
RaWMPC在多个模拟环境中显著优于传统方法，未见场景中碰撞率降低15%，决策透明度提升，模型能主动规避高风险行为。消融实验显示，风险交互训练和自我蒸馏机制是性能提升的关键因素。模型在复杂交通场景中表现出良好的鲁棒性和泛化能力，验证了其在实际应用中的潜力。
Applications
该方法适用于自动驾驶系统的安全决策模块，尤其在复杂未知环境中表现优越。无需专家示范，降低数据采集成本，适合大规模部署。未来可结合真实车辆传感器数据，提升模型在真实环境中的鲁棒性和适应性。
Limitations & Outlook
模型在极端复杂或极端天气条件下仍可能出现风险评估不足的问题。训练过程依赖大量模拟交互，计算成本较高。实际部署中，传感器噪声和环境变化可能影响预测准确性，需进一步增强鲁棒性。未来需结合真实环境数据，优化模型性能。
Plain Language Accessible to non-experts
想象你在厨房做饭，厨师需要不断判断哪些食材新鲜、火候合适。传统方法就像是模仿经验丰富的厨师，只学会了他们的做法，但遇到新菜谱或突发情况时就不知道怎么办。现在，这个新方法像是给厨师配备了一个智能助手，它能提前模拟各种可能的操作后果，主动告诉厨师哪些操作可能会烧焦或不安全。通过不断尝试和学习，助手学会了在不同情况下避开危险，比如避免炒菜时火太大或油溅出来。最终，厨师可以在没有专家指导的情况下，自信地做出安全又美味的菜肴。这就像自主驾驶系统一样，能主动预测和规避潜在风险，确保行车安全。
ELI14 Explained like you're 14
想象你在玩一款赛车游戏，你可以选择不同的路线和操作，但有时候会遇到危险，比如撞到障碍物。以前的赛车游戏只让你模仿高手的操作，学会了他们的路线，但遇到新场景时就不太会应对。现在，这个新方法像是给你的赛车装上了一个聪明的助手，它可以模拟各种可能的操作后果，告诉你哪些操作会很危险，哪些安全。这个助手会自己尝试一些危险的动作，学习它们的后果，然后教你怎么避开危险。这样，即使在你从未见过的复杂赛道上，也能安全快速地跑完全程。这就像让赛车变得更聪明，能自己预测危险，提前做出反应，确保安全。
Glossary
World Model
A predictive model that estimates future environment states based on current observations and actions. It enables simulation of future scenarios for decision-making.
Used to simulate potential outcomes of driving behaviors.
Risk-aware Interaction
A training strategy that systematically exposes the model to risky behaviors to improve its risk prediction and avoidance capabilities.
Core to training the world model without expert labels.
Self-evaluation Distillation
A technique that transfers risk prediction skills from a complex model to a lightweight policy network, enabling efficient inference.
Facilitates real-time low-risk action generation.
Predictive Control
A control approach that uses a model to predict future states and optimize actions over a horizon to achieve desired outcomes.
Fundamental to RaWMPC's decision-making process.
Transformer
A neural network architecture based on self-attention mechanisms, effective for sequence modeling and prediction tasks.
Used for state prediction and traffic event decoding.
Open Questions Unanswered questions from this research
- 1 How to effectively transfer risk prediction capabilities from simulation to real-world scenarios remains open. Handling sensor noise and environmental uncertainties in deployment needs further research.
Applications
Immediate Applications
Autonomous Vehicle Safety Module
Integrate RaWMPC into vehicle control systems to enhance safety in complex, unpredictable environments, reducing collision risks without requiring expert data.
Long-term Vision
Scalable Autonomous Driving Systems
Develop fully autonomous vehicles capable of active risk prediction and avoidance in diverse real-world conditions, accelerating commercial deployment and societal acceptance.
Abstract
With advances in imitation learning (IL) and large-scale driving datasets, end-to-end autonomous driving (E2E-AD) has made great progress recently. Currently, IL-based methods have become a mainstream paradigm: models rely on standard driving behaviors given by experts, and learn to minimize the discrepancy between their actions and expert actions. However, this objective of "only driving like the expert" suffers from limited generalization: when encountering rare or unseen long-tail scenarios outside the distribution of expert demonstrations, models tend to produce unsafe decisions in the absence of prior experience. This raises a fundamental question: Can an E2E-AD system make reliable decisions without any expert action supervision? Motivated by this, we propose a unified framework named Risk-aware World Model Predictive Control (RaWMPC) to address this generalization dilemma through robust control, without reliance on expert demonstrations. Practically, RaWMPC leverages a world model to predict the consequences of multiple candidate actions and selects low-risk actions through explicit risk evaluation. To endow the world model with the ability to predict the outcomes of risky driving behaviors, we design a risk-aware interaction strategy that systematically exposes the world model to hazardous behaviors, making catastrophic outcomes predictable and thus avoidable. Furthermore, to generate low-risk candidate actions at test time, we introduce a self-evaluation distillation method to distill riskavoidance capabilities from the well-trained world model into a generative action proposal network without any expert demonstration. Extensive experiments show that RaWMPC outperforms state-of-the-art methods in both in-distribution and out-of-distribution scenarios, while providing superior decision interpretability.