💻 computer science

Long-Horizon Consistent and Interaction-Aware World Models for Multi-Style End-to-End Driving

The paper proposes **StyleDrive**, a world-model-based framework that addresses temporal inconsistency, interaction modeling, and style adaptability in end-to-end driving through temporal consistency regularization, explicit state disentanglement, and Group Relative Policy Optimization, achieving state-of-the-art performance on the Bench2Drive benchmark and demonstrating successful sim-to-real transfer.

Original authors: Yuxuan Han, Kunyuan Wu, Liyunong Yang, Zilu Wang, Cansen Jiang, Yi Xiao, Liang Hu

Published 2026-09-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuxuan Han, Kunyuan Wu, Liyunong Yang, Zilu Wang, Cansen Jiang, Yi Xiao, Liang Hu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Autonomous vehicles have long relied on a modular approach to driving, breaking the task into separate steps: first seeing the road, then predicting where other cars will go, and finally deciding how to steer. While this method works, it often struggles when the road presents a situation it has never seen before. A newer approach, known as end-to-end learning, attempts to teach a vehicle to map what its sensors see directly to the steering wheel and pedals, much like a human driver learns by experience. To make this learning process safer and more efficient, researchers have begun using "world models." These are internal simulations where the vehicle can imagine future scenarios and practice its reactions without ever leaving the computer. This allows the AI to explore dangerous situations, like a sudden collision, and learn from the outcome without real-world risk. However, these mental simulations often suffer from a specific flaw: as the vehicle imagines further into the future, the picture becomes blurry and inconsistent, and the system often fails to distinguish between a nearby car that matters and a distant one that does not.

A team of researchers has developed a new system called StyleDrive to solve these problems. Their work, tested in a high-fidelity driving simulator and on a real-world vehicle, introduces a way for the AI to maintain a clear and consistent picture of the future while simultaneously learning to drive in different "personalities." The core of their innovation is a method that keeps the vehicle's mental timeline stable. Instead of letting errors pile up as the vehicle imagines seconds into the future, the system constantly checks its current prediction against a history of past states. This acts like a steady hand, ensuring that the imagined road ahead remains physically plausible and coherent, even over long periods. By doing this, the vehicle can plan complex maneuvers, such as merging into fast-moving traffic, with a much higher degree of confidence.

The researchers also addressed how the vehicle pays attention to its surroundings. In previous systems, the AI often treated every other car or pedestrian with the same level of importance, which diluted its focus. The new system explicitly separates the world into two parts: the elements that directly affect the vehicle's safety and movement, and the background elements that do not. This allows the vehicle to concentrate its decision-making power on critical interactions, such as a pedestrian stepping off a curb or a car cutting in, while ignoring irrelevant details. This separation makes the vehicle's choices more interpretable and safer, as it can clearly identify which parts of the environment are driving its actions.

Perhaps the most distinct feature of StyleDrive is its ability to adapt to different driving styles without needing to be retrained from scratch. Most autonomous systems are trained to drive in a single, fixed manner, often leaning toward extreme caution. StyleDrive, however, can learn to drive conservatively, moderately, or aggressively within the same framework. The researchers achieved this by training the system on a group of different driving scenarios simultaneously and comparing the outcomes. Instead of rewarding the vehicle for every small step it takes, the system evaluates entire trips, rewarding the overall success of a journey. This allows the vehicle to learn a spectrum of behaviors, from a cautious driver who waits for perfect conditions to an assertive driver who takes calculated risks, all while maintaining safety.

The results of this approach were significant. When tested on a standard benchmark for autonomous driving, the system achieved a driving score of 88.44 and a success rate of 66.82 percent. These numbers represent a substantial improvement over the previous best methods that used similar world model techniques, which scored significantly lower. The system proved particularly effective in complex scenarios, such as navigating intersections with oncoming traffic or reacting to sudden obstacles. In one test, while other systems either collided with emergency vehicles or drove onto sidewalks, StyleDrive correctly yielded and merged safely. In another, it smoothly overtook stopped traffic without the unsafe maneuvers seen in competing models.

To prove that these skills were not just a result of the simulator, the researchers deployed the system on a real automated vehicle equipped with cameras and laser sensors. In real-world tests, the vehicle successfully handled dynamic situations, such as avoiding oncoming obstacles and stopping for pedestrians who crossed unexpectedly. The system demonstrated that it could transfer the behaviors it learned in the virtual world to the physical world, maintaining its ability to switch between different driving styles depending on the situation. This successful transition from simulation to reality suggests that the system's internal model of the world is robust enough to handle the unpredictability of actual traffic.

The study highlights that the future of autonomous driving may not lie in a single, rigid set of rules, but in flexible systems that can imagine the future clearly and adapt their behavior to the context. By stabilizing long-term predictions and focusing attention on what truly matters, StyleDrive offers a path toward vehicles that are not only safer but also more capable of navigating the diverse and often chaotic reality of human roads. The researchers' work suggests that with the right internal models, machines can learn to drive with a level of nuance and adaptability that was previously out of reach.

Technical Summary: StyleDrive

Problem Statement

End-to-end (E2E) autonomous driving has increasingly adopted world model (WM)-based reinforcement learning (RL) frameworks to improve learning efficiency through imagined rollouts. However, existing WM-based approaches suffer from three fundamental limitations:

Temporal Inconsistency: Long-horizon imagined rollouts often suffer from error accumulation and temporal inconsistency because latent states are generated autoregressively from single-step priors without explicit constraints for long-term physical plausibility.
Inadequate Interaction Modeling: Most models treat surrounding traffic participants uniformly, failing to distinguish between ego-critical agents and less relevant ones. This leads to suboptimal attention allocation and compromised decision-making in complex scenarios.
Inflexible Policy Adaptation: Current policies typically rely on dense, handcrafted reward functions that are sensitive to scaling and struggle to capture multi-step driving intents. Adapting to diverse driving styles (e.g., conservative vs. aggressive) often requires full retraining, limiting behavioral flexibility.

Methodology

The authors propose StyleDrive, a world-model-based RL framework designed to address these limitations through a unified learning paradigm. The framework follows a two-stage training process: World Model Representation Learning and Policy Learning.

1. World Model Representation Learning

The framework employs a Recurrent State Space Model (RSSM) enhanced with specific architectural innovations:

Explicit State Disentanglement: The latent state is decoupled into a controllable ego-state ( $z^ego_t$ ) and an interactive environment state ( $z^{env}_t$ ). The interactive state is further decomposed into ego-relevant and ego-irrelevant components. This is achieved via a two-branch structure using dedicated Gated Recurrent Units (GRUs).
Temporal Consistency Regularization: To stabilize long-horizon rollouts, the model integrates historical latent context via gated cross-attention. Instead of relying solely on a single-step prior, the model constructs a temporal context sequence from past prior states and uses a learnable gating mechanism to fuse this history with the current prior, mitigating error accumulation.
Semantic Dependency: The relevance of interactive states is refined using a semantic dependency score based on geometric proximity and semantic interaction, suppressing low-dependency regions to enhance interpretability and safety.

2. Policy Learning

StyleDrive introduces a multi-style policy optimization strategy:

Multi-Style Configuration: Driving styles (Conservative, Moderate, Aggressive) are defined by semantically partitioning key reward weights (e.g., collision penalty, waypoint reward). The policy is conditioned on these style configurations ( $\alpha(n)$ ).
Group Relative Policy Optimization (GRPO): The framework replaces standard per-step reward optimization with trajectory-wise relative advantages. By deploying the agent in parallel environments, the method computes group-normalized returns. This intra-group relative ranking encourages high-return trajectories without requiring per-step advantage estimation or a value function, enabling the joint training of diverse driving styles without retraining or reward rescaling.

Key Contributions

The paper identifies four primary contributions:

Temporal Consistency Regularization: A mechanism integrating historical latent context via gated cross-attention to stabilize long-horizon imagined rollouts and reduce error compounding.
Explicit State Disentanglement: A module that separates ego-relevant from ego-irrelevant interactive states, enabling interpretable and risk-aware decision-making.
Group Relative Policy Optimization (GRPO): A framework that replaces per-step rewards with trajectory-wise relative preferences, facilitating the joint training of multiple driving styles.
Sim-to-Real Deployment: A practical demonstration of the framework's transfer capability on a real automated guided vehicle (AGV) platform, showing robust performance in dynamic driving scenarios.

Experimental Results

The authors evaluated StyleDrive on the Bench2Drive closed-loop driving benchmark and deployed it on a real-world AGV platform.

Quantitative Performance: On the Bench2Drive benchmark, the multi-modality variant of StyleDrive achieved a Driving Score (DS) of 88.44 and a Success Rate (SR) of 66.82%. These results represent an improvement of +17.08 in DS and +16.58% in SR over the previous best world model-based method (Epona). In the camera-only setting, it also achieved the best performance among single-modality models (DS: 79.78, SR: 57.91%).
Multi-Ability: The model demonstrated robustness across five isolated driving skills (Merging, Overtaking, Emergency Brake, Giving Way, Traffic Sign), achieving a new state-of-the-art mean score of 64.71 in the multi-modality setting.
Ablation Studies:
- Disentanglement: Explicit decomposition of the environment state into ego-relevant and ego-irrelevant components via cross-attention yielded significant gains (+8.01 DS, +9.20% SR) over variance-based or simple ego-environment splits.
- Policy Optimization: Integrating multi-style policies with GRPO fine-tuning resulted in the highest performance across all metrics, with a Diversity score of 0.524 (a 68% increase over multi-style PPO without GRPO).
- Temporal Consistency: The regularization significantly improved closed-loop performance and open-loop rollout fidelity. In open-loop tests, the model maintained lower Fréchet Inception Distance (FID) and Fréchet Video Distance (FVD) scores compared to baselines, with the performance gap widening as the prediction horizon increased.
Real-World Deployment: The system was successfully deployed on an AGV platform equipped with LiDAR and cameras. It demonstrated successful sim-to-real transfer in scenarios involving oncoming obstacle avoidance, static obstacle avoidance, and emergency responses to jaywalking pedestrians, adapting its behavior to both aggressive and conservative styles.

Significance

The paper claims that StyleDrive addresses critical gaps in existing world model-based driving frameworks. By enforcing long-horizon consistency, explicitly disentangling interactive states, and enabling multi-style optimization without retraining, the framework offers a more robust, interpretable, and flexible solution for E2E autonomous driving. The successful sim-to-real transfer validates the practical applicability of the proposed approach in dynamic, real-world traffic environments.