[ArXiv 2026] RaWMPC: Beyond Mimicry—Breaking the Generalization Dilemma of Autonomous Driving via Risk-Aware World Models

Executive Summary

TL;DR: Researchers from the University of Trento and Sun Yat-sen University have unveiled RaWMPC, the first end-to-end autonomous driving (E2E-AD) framework that achieves SOTA performance without requiring any expert action supervision. By shifting the paradigm from "driving like an expert" to "proactively avoiding predicted risks," RaWMPC solves the generalization issues inherent in imitation learning, particularly in rare and hazardous "long-tail" scenarios.

Academic Position: This work moves beyond the standard Imitation Learning (IL) and Model-Based Reinforcement Learning (MBRL) paradigms. It introduces a pure Predictive Control approach where a world model is specifically trained to understand "what goes wrong" through a unique risk-aware interaction strategy.

1. The Generalization Dilemma: Why Mimicry Fails

The dominant paradigm in E2E-AD is Imitation Learning (IL). While successful in standard conditions, IL suffers from a fundamental flaw: Experts never crash.

Because the training data consists only of safe, smooth expert trajectories, the model never learns the consequences of bad decisions. When faced with an unfamiliar scenario (e.g., a sudden cut-in during heavy rain), the model doesn't know what a collision "looks" like in its latent space, leading to unpredictable and often fatal behaviors.

2. Methodology: RaWMPC Architecture

RaWMPC addresses this by treating the driving task as a World Model Predictive Control problem.

The World Model & Semantic Decoding

Instead of outputting a single action, RaWMPC takes $N$ candidate action sequences and "imagines" their consequences using a Transformer-based world model. The system doesn't just predict pixels; it uses Semantic-Guided Decoding to predict:

1. Semantic Segmentation: High-level visual understanding.
2. Traffic Events: Probabilities of collision, off-road driving, or traffic sign violations.
3. Future Ego-states: Predicted velocity and position.

Risk-Aware Interaction Strategy

To make the world model "risk-intelligent," the authors propose a training scheme with three modes:

• Good Mode: Samples from low-cost (safe) candidates.
• Rand Mode: Broad exploration.
• Bad Mode: Crucially, this mode intentionally selects high-risk actions to interact with the simulator. This allows the world model to observe and learn from "catastrophic outcomes," making them predictable and thus avoidable during test time.

3. Self-Evaluation Distillation

To solve the high computational cost of sampling actions at test time, RaWMPC uses Self-Evaluation Distillation. The world model acts as a "teacher" that scores millions of random action samples. The best (positive) and worst (negative) samples are used to train a Conditional VAE (cVAE) action proposer through InfoNCE Contrastive Learning. This ensures the generative policy naturally proposes low-risk candidates for the final MPC refinement.

4. Experimental Results: Breaking the SOTA

RaWMPC was tested on Bench2Drive (closed-loop) and NAVSIM (real-world data).

Key Performance Metrics:

• Bench2Drive: Achieved 88.31 Driving Score (DS), outperforming HiP-AD (86.77) and SimLingo (85.94).
• Zero-Expert Achievement: Even with 0% expert data, RaWMPC scored 87.34 DS, proving that environment interaction can replace costly human labeling.
• Robustness: In a "Sunny-to-Rainy" domain shift test, RaWMPC maintained a DS of 41.36, while strong IL baselines like LAW dropped to 23.58.

Visualization of Predictive Control

In complex scenarios (e.g., a pedestrian crossing while a car is merging), RaWMPC demonstrates "intuitive" decision-making. By comparing "stop," "swerve," and "accelerate," the model selects the maneuver that maximizes progress while keeping the predicted collision probability at zero.

5. Critical Analysis & Takeaways

Strengths:

• Interpretability: Unlike "black-box" IL models, RaWMPC's decisions are explainable via its predicted costs and future semantic rollouts.
• Generalization: By learning physical constraints and collision consequences, it handles out-of-distribution weather and traffic better than imitation-based models.

Limitations:

• Sim-to-Real: The "Bad Mode" interaction requires a high-fidelity simulator (CARLA). Transitioning this risk-aware training to the real world remains a challenge due to safety risks during exploration.
• Horizon accumulated error: As shown in the ablation study, looking too far ahead (H>15) increases uncertainty and degrades ranking accuracy.

Conclusion: RaWMPC represents a significant milestone in E2E autonomous driving. It proves that a "knowledge-driven" world model, trained to understand danger, is more robust than a "data-driven" model trained merely to follow expert paths.