World Engine: Towards the Era of Post-Training for Autonomous Driving

Abstract: Autonomous vehicles must operate safely in the real world, where errors can have severe consequences. Although modern end-to-end driving policies excel in routine scenarios, their reliability is limited by the scarcity of safety-critical ``long-tail'' events in real driving datasets. These rare interactions define the practical safety boundary of the learned policy, yet they are difficult to collect at scale in the real world. Here we show that this fundamental limitation can be addressed by post-training pre-trained driving models on synthesized high-stakes interactions. We introduce World Engine, a generative framework that reconstructs high-fidelity interactive environments from real-world logs and systematically extrapolates them into realistic safety-critical variations. This paradigm enables reinforcement-based post-training to align policies with safety constraints, circumventing the physical risks inherent in real-world exploration. On a public benchmark built on nuPlan, World Engine substantially reduces failures in rare safety-critical scenarios and yields significantly larger gains than scaling pre-training data alone. Furthermore, when deployed on a production-scale autonomous driving system, the resulting policy reduces simulated collisions and demonstrates measurable improvements in on-road testing, showing that post-training on synthesized, safety-critical interactions offers a scalable and effective pathway to safer autonomous driving. The full codebase suite, including training, is released to the public.

• The paper introduces the World Engine framework that redefines policy improvement as a post-training paradigm by unifying failure discovery, photorealistic simulation, behavior modeling, and reinforcement learning.
• It demonstrates significant improvements in rare-event robustness, achieving up to an 88.9% closed-loop success rate and a 45.5% reduction in rare cut-in collisions.
• The approach offers enhanced data efficiency, delivering safety gains comparable to a 10x increase in passive data while constraining policy drift through KL-regularized updates.

World Engine: A Post-Training Paradigm for Safety-Critical Autonomous Driving

Introduction and Problem Statement

Modern end-to-end autonomous driving systems demonstrate strong competency in common scenarios but exhibit notable brittleness in the "long tail" of rare, safety-critical events—situations such as sudden pedestrian crossings and aggressive vehicle cut-ins that, while seldom encountered, disproportionately define system safety boundaries. Accumulating vast amounts of mundane driving data yields diminishing returns in these edge regimes due to their natural scarcity and the ethical and logistical impossibility of mining safety-critical interactions at scale from the real world. This structural paradox fundamentally limits robustness and reliable policy deployment.

World Engine Framework

World Engine redefines safety-critical policy improvement as a post-training problem, in direct analogy to recent advances in LLM post-training. Rather than exclusively relying on passive log accumulation, World Engine unifies failure-grounded discovery, controllable simulation, interactive behaviour modeling, and behaviour-regularized RL in a closed-loop pipeline to specifically densify the data distribution over sparse long-tail events.

The pipeline consists of four main phases:

1. Base Policy Pre-Training and Failure Discovery: An end-to-end policy is first trained by imitation learning on canonical fleet-scale data. This policy then processes real-world logs through open-loop rollouts to automatically surface failure-prone scenarios, establishing a targeted subset for augmentation.
2. Photorealistic Interactive Simulation: Discovered long-tail scenes are reconstructed via a 3D Gaussian Splatting pipeline, enabling high-fidelity, compositional scene manipulation. This allows for controlled variation of both static infrastructure and dynamic agents, supporting novel sensor observations under arbitrary agent configurations.
3. Behaviour World Modeling: A learned, diffusion-based world model generates realistic, multi-agent trajectories for surrounding agents responsive to ego policy intent. This supports both stochastic diversity and goal-oriented/adversarial scenario synthesis, going beyond log-replay or rule-based models to expose the ego policy to a rich, reactive interaction spectrum.
4. Reinforcement-Based Policy Post-Training: The agent is refined via RL on simulated long-tail rollouts and real-world logs, optimizing a shaped reward that integrates safety, efficiency, and comfort objectives while regularizing policy drift via a KL term to constrain deviation from the pre-trained policy. Experience sampling emphasizes hard, informative frames to target learning to critical failure regimes.

This closed-loop data generation-training-evaluation cycle robustly addresses causal feedback loops and rare event coverage, moving beyond passive expansion of static datasets.

Empirical Results: Simulation and Real-World Validation

World Engine is demonstrated across both public academic (nuPlan) and industrial-scale (Huawei ADS) platforms. Key empirical findings include:

• Long-Tail Safety-Critical Robustness: On nuPlan rare-event benchmarks, World Engine post-training improves rare scenario closed-loop success rates from 73.7% to 88.9% and PDMS* from 60.98 to 70.12, while maintaining or improving open-loop performance on common cases.
• Data Efficiency: Post-training on synthesized safety-critical scenarios achieves higher rare-case robustness than doubling the pre-training data scale; extrapolated, World Engine achieves safety improvements comparable to a 10x increase in passively collected data.
• Production-Scale Deployment: On Huawei ADS, a policy base-trained on 80,000 hours of logs achieves a further 45.5% reduction in rare cut-in collisions and consistently lowers failure rates across all major safety metrics after World Engine post-training. Closed-loop hardware-in-the-loop and >200 km real-world on-road tests confirm zero disengagements for post-trained policies and demonstrably improved handling of cut-in and occluded pedestrian events compared to base policies.

These findings argue strongly for the practical scalability, safety impact, and transferability of post-training on synthesized rare events.

Technical Innovations

World Engine introduces several critical advances:

• Grounded Failure Discovery: Failure modes are discovered by deploying the base policy on real-world logs, ensuring that synthetic augmentation is physically plausible and policy-aligned.
• 3DGS-based Neural Scene Reconstruction: High-fidelity, real-time rendering retains realism and supports interactive closed-loop rollouts far from original logged trajectories.
• Diffusion-Based Behaviour World Model: Enables diverse, controllable, and human-aligned multi-agent interaction generation, supporting explicit adversarial scenario synthesis and behaviour regularization during scene sampling.
• Behaviour-Regularized RL Post-Training: Mixture-of-experience distributions and KL-constrained updates prevent catastrophic forgetting of routine competence while emphasizing safety-critical gains.
• Efficient and Scalable Rollout Pipeline: Achieves high-throughput simulation compatible with large-scale post-training and policy evaluation.

Implications and Future Directions

Practically, World Engine provides a scalable, data-efficient route to addressing the long-tail safety problem in autonomous driving, reducing the dependence on massive, costly, and largely redundant log collection. Theoretically, it shifts the paradigm from distributional coverage via brute-force data expansion to targeted data densification and post-training in rare regimes guided by generative modeling and reinforcement learning.

In its current form, the method is limited by its reliance on failure types present in historical logs and by the sim-to-real fidelity of 3DGS and behaviour models, particularly as agent behaviour departs from logged distributions. Future research should target:

• Extending failure discovery via adversarial or procedural scenario generation;
• Improving neural world modeling (e.g., foundation video models) for broader transfer and realism;
• Iterative multi-round post-training for persistent correction of emerging rare failure modes;
• Generalization to other Physical AI domains—robotic manipulation, surgical robotics, etc.—that similarly suffer from under-representation of consequential rare events.

Conclusion

World Engine demonstrates that post-training autonomous driving policies on synthesized, physically grounded safety-critical scenarios delivers substantial improvements in rare-case robustness, with strong data efficiency and no compromise to standard-case proficiency. The approach forms a blue-print for reliably safe physical AI systems: discover failure, reconstruct, synthesize, and reinforce, with the entire pipeline grounded in real-world empirical evidence and extensible through improvements in generative simulation and world modeling.

(2606.19836)