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

1 Introduction

• Rare events such as abrupt pedestrian crossings, aggressive cut-ins, and complex interactions remain statistically sparse, although they define the practical safety boundary of learned policies.Society cannot safely collect these events at scale, leaving valuable learning signals systematically missing from naturalistic datasets.Source: §1 · pp. 1–2
• The paper frames passive data scaling as insufficient for rare cases, reporting diminishing returns from added uneventful logs.The authors describe this as a long-tail data problem that motivates learning beyond passive data accumulation.Source: §1 · p. 2
• World Engine discovers failure-prone scenarios in real logs, reconstructs them as high-fidelity interactive worlds, and applies reinforcement learning post-training without additional real-world exploration risk.The approach draws on post-training work that uses synthesized reasoning and reinforcement learning to address sparse high-value data.Source: §1 · p. 2
• World Engine improves closed-loop success on nuPlan long-tail cases, with gains substantially larger than doubling pre-training data and extrapolated to remain competitive with ∼10× more data.The comparison with ∼10× more pre-training data is an extrapolation from the scaling trend.Source: §1 · p. 2
• Up to 45.5% reduction in collision rate followed post-training on a production-scale system; a 200-kilometre on-road test recorded zero disengagements and improved safety in safety-critical scenarios.The collision-rate reduction was measured in closed-loop simulation across over 10,000 scenarios.Source: §1 · p. 2

2 Overview of World Engine

• The four-stage pipeline pre-trains an agent, discovers its failure-prone logged events, reconstructs and varies them in simulation, then applies reinforcement post-training.World Engine shifts the closed-loop training distribution toward long-tail scenarios beyond what real-world collection alone provides.Source: §2 · p. 3, Figure 1 · p. 3

2.1 Pre-training and Long-tail Event Discovery

• A pre-trained agent flags logged scenarios where its planned trajectory collides with objects or leaves the road, selecting physically grounded cases aligned with its competence boundary.The selection uses a lightweight simulator with 3D bounding boxes and HD maps; grounding events in real sensor data and traffic configurations supports physical plausibility.Source: §2.1 · p. 4

2.2 Simulation Engine

• The 3DGS simulation engine separates static backgrounds from dynamic agents for object-level editing, while free-viewpoint rendering supports novel closed-loop trajectories and thousands of rollouts per scenario.Each scene element is represented by 3D Gaussians fitted to multi-view log observations.Source: §2.2 · p. 4

2.3 Behaviour World Model

• A diffusion behaviour model generates diverse, ego-responsive traffic trajectories; goal conditioning and optimization guidance steer samples, while log replay and IDM provide alternative traffic models.A single long-tail scenario can be expanded into hundreds of variations.Source: §2.3 · pp. 4–5

2.4 Reinforcement Post-training

• Behavior-regularized reinforcement learning combines logged and simulated experience, prioritizes failure or near-failure frames, and uses a KL constraint to improve safety-critical behavior while preserving common-driving competence.Dense rewards guide safety, efficiency, and comfort objectives.Source: §2.4 · pp. 5–6

3.1 Implementation of World Engine

• The nuPlan reference implementation and industrial deployment preserve the same conceptual pipeline and learning objectives.The paper uses the nuPlan implementation for controlled ablations and quantitative analysis.Source: §3.1 · p. 6

3.2 Simulation Engine

• The simulation engine pairs 3D Gaussian scene reconstruction, including separately reconstructed dynamic objects, with controllable rendering that manipulates objects for real-time closed-loop observation generation.The scene graph separates static background structure from dynamic traffic participants.Source: §3.2 · pp. 6–7

3.3 Behaviour World Model

• The behaviour world model uses diffusion to synthesize stochastic and adversarial traffic, representing agent behaviours and map features as tokenized sequence states for simultaneous agent-future prediction.Classifier guidance adjusts generated trajectories to avoid collisions, smooth paths, and pull agents toward nearby lanes.Source: §3.3 · pp. 8–9

3.4 Reinforcement Post-training

• Reinforcement post-training combines curated logged and simulated experience with safety-, progress-, and comfort-related rewards and hard-example mining, while regularizing toward the pretrained expert.Candidate samples include rare or safety-critical interactions and are checked for physical plausibility.Source: §3.4 · pp. 9–10

3.5 Base End-to-End Driving Model

• The base planner uses BEVFormer multi-camera encoding, UniAD-style tracking and map-segmentation supervision, and a scoring decoder that selects among predefined trajectories.The selected trajectory is conditioned on the bird’s-eye-view feature representation.Source: §3.5 · p. 10

4.1 Implementation Details

• NuPlan post-training used 31,508 generated frames from 5,340 mined long-tail scenes; the production system mixed 1.0 million World Engine clips with 5.0 million common-driving clips.The production pre-training corpus comprised approximately 80,000 hours of driving data.Source: §4.1 · pp. 10–11

4.2 Safety-critical Closed-loop Simulation

• The nuPlan benchmark evaluates disjoint, collision-centric four-second rare-event clips with open-loop PDM Score and closed-loop success defined by avoiding collision or off-road infraction.The rare-event evaluation subset contains 288 scenarios.Source: §4.1 · p. 10, §4.2 · p. 11
• On a 50k-scene base model, World Engine post-training exceeded the pre-training scaling trend across common open-loop, rare open-loop, and rare closed-loop measures.The paper estimates that comparable closed-loop gains from pre-training alone would require approximately an order-of-magnitude more real-world data.Source: §4.2 · p. 11
• Rare logged scenes improved rare open-loop PDMS but yielded limited closed-loop gains; reactive rare rollouts and behaviour-model traffic augmentation produced stronger rare closed-loop results.The full pipeline reached rare closed-loop PDMS* of 70.12, versus 67.33 without the behaviour world model; synthetic replays had the lowest ego progress in the comparison.Source: §4.2 · pp. 11–12, Figure 2 · p. 12

4.3 Production-scale Driving Validation

• For production-scale validation, failure-prone scenarios were reconstructed with 3DGS-based rendering, traffic interactions were augmented with a behaviour world model, and the policy was refined through reinforcement post-training.Evaluation used asynchronous hardware-in-the-loop closed-loop simulations on production on-board hardware.Source: §4.3 · p. 13
• World Engine achieved rare closed-loop SR of 88.89% and PDMS* of 70.12 on nuPlan, while production simulations reduced cut-in collisions by 45.5%.Across three on-road runs totaling 200 km, the post-trained model had zero disengagements; the base model triggered one safety-critical intervention.Source: §4.3 · pp. 13–14, Table 1 · p. 14

Results

• World Engine reduces safety-critical benchmark failure rates by up to 45.5% in cut-in tests while maintaining performance on common cases.Its safety gains can match those of approximately 10× more pre-training data, based on extrapolating the scaling curve.Source: §5 · p. 14

Conclusion

• World Engine shifts long-tail learning from passive collection to discovering failures in real logs, reconstructing interactive worlds, generating counterfactual variations, and applying reinforcement learning.The authors describe this targeted approach as a more cost-effective path that can shorten the iteration cycle for improving autonomous driving systems.Source: §5 · pp. 14–15

Production evaluation

• Figures 4–6 illustrate production-level closed-loop safety metrics, Shanghai on-road routes and interactions, and nighttime scenarios.Figure 4 compares the ADS base and post-trained models; Figures 5 and 6 show representative on-road and nighttime cases.Source: Figure 4 · p. 15, Figure 5 · p. 16, Figure 6 · p. 17

Limitations

• The framework cannot discover scenario types absent from collected logs, and its simulation fidelity is limited when rollouts depart substantially from recorded trajectories or involve some road users.The behavior world model is particularly limited for pedestrians, cyclists, and unstructured road users; post-training rewards are manually defined and may miss nuances of human driving preferences.Source: §5 · pp. 15–16

Scalability and iterative refinement

• With the 58.3M-parameter model, multiple post-training rounds tend to destabilize the policy, leaving stable iterative refinement an open question.The authors suggest larger models or more sophisticated regularization as possible ways to support stable multi-round training.Source: §5 · pp. 16, 18

Results

• Even with fewer identified long-tail events, post-training yields meaningful safety improvements as the base model grows stronger.The Huawei ADS base model was trained on over 80,000 hours of data, and had a substantially smaller proportion of scenarios identified as long-tail than the academic setting.Source: §5 · p. 18

Towards general Physical AI

• The authors propose the discover–world-modelling–post-train pipeline as a template for other Physical AI domains, with implementation choices varying by application.They name robot manipulation, legged locomotion, and surgical robots as examples, and point to video world models or physics-based simulation as possible alternatives to driving-specific components.Source: §5 · p. 18

A.2 Learning Pipeline in Autonomous Driving

• End-to-end policies are refined through multiple post-training approaches, but existing methods can be constrained by instability, simulator bias, or limited interaction.Prior approaches include simulation-based reinforcement learning, residual or adapter updates, expert distillation, test-time adaptation, and model-based reinforcement learning.Source: §A.2 · pp. 25–26

A.1 Data Engine for Autonomous Systems

• Data engines and simulation have expanded datasets and supported validation, but generated scenarios are typically used to test robustness rather than directly guide learning.Existing systems often operate in open-loop settings, without capturing the causal effects of policy actions or coupling behavior generation to evolving policy weaknesses.Source: §A.1 · p. 25, §A.3 · p. 26

A.3 Comparisons with Related Work

• World Engine couples policy rollouts in photorealistic interactive environments with scenario curation and reinforcement post-training, targeting informative long-tail cases.The framework combines neural scene reconstruction with behavior-level world modelling to generate diverse and counterfactual interactions.Source: §A.3 · p. 26

B.1 Behaviour World Model

• The behavior world model uses a diffusion model with individual noise levels to generate realistic, reactive, and controllable traffic behavior.This model underpins the World Engine system.Source: §B.1 · p. 26

B.1.1 Scene Vectorized Representation

• Scenes use an Agent Tensor for dynamic entities and a Map Tensor for static features, with future scenes generated from history, the current time step, and the map.Agent attributes include positions, headings, velocities, and dimensions; map features represent lanes.Source: §B.1.1 · p. 26

B.1.2 Decoupled Noise Modelling

• Decoupled triaxial mask modelling assigns independent noise levels across agents, time steps, and denoising steps to model reactive traffic efficiently.During sampling, low-noise history and goals condition generation of the other scene tokens.Source: §B.1.2 · p. 27

B.1.3 Classifier Guidance for Human-behaviour Alignment

• Classifier guidance enforces collision avoidance, smooth acceleration, and on-road driving by separating overlapping vehicles, smoothing trajectories, and pulling vehicles toward lane centrelines.The guidance adjusts agent trajectories to align generated scenes with human-behaviour rubrics.Source: §B.1.3 · p. 27

B.1.4 Inference Mode

• Inference generates safety-critical scenes through Scenario Copy, which perturbs existing agents’ goals, and Intent Attack, which redirects a selected neighbour toward the ego vehicle’s goal.The world model generates interactive trajectories conditioned on the perturbed or adversarial goals.Source: §B.1.4 · p. 28
• Generated scenarios are refined and retained only after collision and drivable-area checks, while iterative generation stops at a target scenario count or repeated-failure threshold.The process uses an LQR-based tracker before validation.Source: §B.1.4 · p. 28

B.2 Reinforcement Post-training

• The post-training framework integrates curated corner-case sampling and verifiable reward shaping with well-defined optimization objectives for scalable large-rollout training.It is designed to support continuous and active improvement of driving performance.Source: §B.2 · p. 28

B.2.1 Data Sampler

• Corner-case curation samples trajectories from fleet logs, user data, risk- or uncertainty-based mining, and counterfactual rollouts, organizing them by source and semantic category.Curriculum weighting controls how strongly trajectories influence policy updates, alongside source- and category-level sampling.Source: §B.2.1 · p. 29
• Verifiable reward modules combine explicit safety gates with weighted objectives, allowing safety-critical violations to sharply suppress the overall reward.The gates are described as enforcing nonnegotiable safety requirements.Source: §B.2.1 · pp. 29–30, Table S1 · p. 29
• Progress and time-to-collision receive equal, highest weight, while comfort has lower weight so necessary evasive manoeuvres are not penalized.Equal progress and time-to-collision weighting helped avoid policies that halt indefinitely or accelerate through near-miss events.Source: §B.2.1 · p. 30, Table S2 · p. 30

B.2.2 Rollout Platform

• The rollout platform uses high-throughput rendering and behaviour modelling to generate diverse, rare interactions, monitors safety and performance, and rolls back detected degradation.Deployment expands progressively while supporting a continuous cycle of data collection, curation, and policy improvement.Source: §B.2.2 · p. 30

B.3 Simulation Engine: Closed-loop Platform

• The closed-loop testing platform supports large-scale log-based evaluation with multi-traversal neural rendering, parallel rollouts, and configurable traffic interaction fidelity.It is designed to assess safety, robustness, and compounding-error effects over long horizons.Source: §B.3 · p. 30, Figure S1 · p. 31

B.3.1 Platform Setup

• Each rollout starts from logged states and cached rendered scenes; the policy supplies waypoints or controls, while traffic may be replayed, IDM-controlled, or world-model-generated.The described closed-loop evaluation uses IDM-controlled traffic and terminates on hard safety or route constraints, successful completion, or the time horizon.Source: §B.3.1 · pp. 31–32
• Distributed simulation and inference workers exchange batched observations and actions asynchronously to support scalable closed-loop evaluation.The modular components expose standardized execution hooks.Source: §B.3.1 · p. 32

B.3.2 Testing Metrics

• Open-loop evaluation uses PDMS for short-horizon, non-reactive driving, aggregating safety, legality, efficiency, and comfort measures.Its formula combines no-at-fault collisions, drivable-area compliance, time-to-collision, ego progress, and comfort.Source: §B.3.2 · p. 32
• PDMS scores no-at-fault collisions, drivable-area compliance, time-to-collision, ego progress, and comfort through discrete or bounded component measures.The components assess responsible collisions, road-boundary adherence, near-miss risk, route progress, and human-calibrated kinematic thresholds.Source: §B.3.2 · p. 32
• Closed-loop Success Rate averages rollouts that complete without collision or drivable-area violation, reflecting safety-critical feasibility.The metric is episode-level because closed-loop interactions compound over time.Source: §B.3.2 · p. 33
• PDMS* evaluates posterior closed-loop ego and traffic trajectories using human-driven progress references and reactive traffic behaviour.These changes distinguish it from standard PDMS, which uses a PDM-Closed progress upper bound and assumes non-reactive background agents.Source: §B.3.2 · p. 33

• 576 scenarios were processed in 25.48 minutes, using 6.63 GPU-hours at 11.59 scenarios per minute.Source: Table S3 · p. 33, §B.3.3 · p. 33
• NR and R stages run in parallel, so total wall time equals the larger stage time, max(24.22, 25.48).Source: §B.3.3 · p. 33
• Per-step planning takes ∼3,800 ms for model inference, ∼490 ms for 3DGS sensor rendering, and ∼215–350 ms for simulator state advancement.Source: §B.3.3 · p. 33
• Reactive evaluation adds 5.5% scenario-level overhead over non-reactive evaluation, primarily from simulator state advancement for reactive traffic generation.Source: §B.3.3 · p. 33

• Reconstruction stability is assessed across 12,862 navtrain assets, with image quality measured by PSNR and SSIM and LiDAR depth quality by δ1.δ1 is the fraction of predicted depth within a factor of 1.25 of ground-truth depth.Source: §B.4 · p. 34, Figure S2 · p. 34
• The simulation engine renders realistic multi-view observations for simulated ego positions that deviate from the logged trajectory, supporting closed-loop rollouts.Source: §B.4 · p. 34, Figure S3 · p. 35