Introduction
Autonomous driving systems based on behavior cloning often struggle with the "generalization gap." A model trained exclusively in clear weather may fail catastrophically when encountering rain or night conditions. This failure stems from the model's inability to distinguish between causal factors (e.g., lane markings) and nuisance factors (e.g., lens flares or sky color). The MoDE (Mixture of Domain-Specific Experts) framework addresses this by disentangling visual representations into domain-general, domain-specific, and dynamic-object features, using a specialized mixture-of-experts architecture to maintain performance across diverse environments.
The Generalization Challenge
Standard behavior cloning models often conflate environmental context with driving logic. For instance, a model might learn to associate the specific hue of a "Clear Noon" sky with the act of driving straight. When tested in a "Rainy Sunset," the change in sky color and road reflections acts as a distribution shift that confuses the agent. Furthermore, dynamic objects like pedestrians and other vehicles introduce high variance that standard encoders struggle to represent reliably without explicit supervision. To bridge this gap, the authors propose forcing the model to "disentangle" these factors, effectively teaching it to ignore environmental noise while focusing on the road and traffic.
The MoDE Framework
The MoDE architecture utilizes a two-stage learning process centered on feature disentanglement and conditional execution.
1. Feature Disentanglement
The system splits the latent representation of input images into three disjoint components:
- Domain-General ($f^z$): Captures static road geometry (lanes, curbs) that remains constant regardless of weather.
- Domain-Specific ($f^{\tilde{z}}$): Captures environmental factors like lighting and weather conditions.
- Dynamic-Object ($f^{\hat{z}}$): Represents transient elements like vehicles and pedestrians.
This separation is achieved using a Cycle-Consistent VAE and a Mutual Information (MI) estimator. The MI estimator maximizes the information shared between dynamic-object features and the raw image while minimizing overlap with other latent spaces, ensuring that moving objects are cleanly isolated from the background.
2. Mixture of Domain-Specific Experts
Instead of a monolithic policy, MoDE employs multiple "experts." A gating network analyzes the domain-specific features (weather/lighting) to assign weights ($g_1, ..., g_n$) to different experts. The final control command $\hat{a}$ (steering, throttle, brake) is a weighted sum of the experts' outputs, which are themselves conditioned on the domain-general and dynamic-object features.
3. Kinematic Regularization
To ensure the perception network learns physically grounded representations, the authors include a Location Offset subtask. This task requires the model to predict the vehicle's future position based on current speed and steering, acting as a structural constraint on the learned features.
Experimental Evaluation
The framework was evaluated using the CARLA simulator (v0.8.4) on two rigorous benchmarks: NoCrash (varying traffic density) and AnyWeather (unseen weather conditions like "Soft Rain Sunset").
Key Results
- Superior Generalization: In the "New Town & Weather" benchmark—the most difficult test of generalization—MoDE achieved a 68% success rate in regular traffic, significantly outperforming baseline models like CILRS (56%) and FASNet (66%).
- Robustness to Unseen Weather: On the AnyWeather benchmark, MoDE maintained higher success rates across navigation tasks compared to competitors, proving the effectiveness of the gating mechanism in handling novel environmental shifts.
- Ablation Insights: Removing the MI estimator resulted in the most significant performance degradation, highlighting that the ability to disentangle dynamic objects is critical for safety in complex traffic.
Conclusion
MoDE demonstrates that structured feature disentanglement is a powerful tool for autonomous driving. By separating the permanent rules of the road from fleeting environmental conditions, the model achieves a level of robustness that eludes standard end-to-end approaches. Future work will focus on improving performance in extremely dense traffic and expanding the diversity of expert training to handle more complex maneuvers beyond standard navigation.
Summary updated