A new end-to-end autonomous driving system gives operators the ability to dial in a specific driving style—from conservative lane-cuddling to aggressive lane-changing—without retraining the model. Researchers at Changchun University developed D³-MoE, which uses a dual mixture-of-experts approach with separate lateral and longitudinal expert networks, all steered by a diffusion-based trajectory generator that translates your style preference directly into motion.
What the Researchers Built
D³-MoE is an end-to-end autonomous driving model that takes multi-sensor inputs—six cameras, LiDAR, HD maps—and directly outputs a 4-second trajectory composed of eight waypoints. What sets it apart is its ability to control the style of the generated trajectory at inference time using a simple discrete style label: aggressive, conservative, or neutral.
To achieve this, the architecture uses two separate mixture-of-experts (MoE) modules—one for lateral behavior (turning) and one for longitudinal behavior (speed). The lateral MoE has five experts: sharp left, left, straight, right, sharp right. The longitudinal MoE has three: braking, cruising, accelerating. Two independent routers select which expert to activate based on the input data and the desired style. The chosen experts then guide a conditional diffusion model that denoises a random trajectory into a smooth, style-consistent path.
The training data was augmented with synthetically aggressive and conservative trajectories to teach the model the full spectrum of driving personalities. The result is a single model that can be tuned on the fly without additional fine-tuning.
Key Results
The model was trained and evaluated on the NAVSIM dataset, which contains driving scenes from multiple US cities. D³-MoE was compared against several baselines including vanilla diffusion, unconditional MoE, and end-to-end planners like UniAD and VAD.
The primary metric is the Predictive Driving Model Score (PDMS) and its extended variant EPDMS. The paper reports that D³-MoE achieves the highest PDMS among all methods while also producing the most style-consistent trajectories. Specifically:
- Style controllability: When given an aggressive style label, the model-generated trajectories displayed higher speed profiles and more frequent lane changes. Under a conservative label, trajectories showed lower speeds and tighter lane centering.
- Safety metrics: D³-MoE maintained near-zero collision rates and high time-to-collision values across all styles, with the conservative style achieving a 100% no-collision rate on the validation set.
- Comfort: The aggressive mode produced slightly jerkier movements (comfort score drop of ~5%), but still within acceptable passenger comfort thresholds.
- Generalization: The model retained style-consistent behavior even in unseen road geometries and traffic conditions.
The extended EPDMS metric, which includes lane keeping and traffic light compliance, showed that all style variants comply with basic traffic rules, with the conservative mode rarely requiring hard braking.
How It Works
D³-MoE processes a 2-second history of multi-view images, LiDAR point clouds, and map data to produce an 8-waypoint trajectory spanning the next 4 seconds. The core innovation is the dual disentangled mixture-of-experts.
Dual Decoupled MoE: Rather than using a single MoE that must learn all aspects of driving, the architecture separates lateral and longitudinal control. Each has its own set of experts and its own router (a lightweight neural network). The routers receive both the sensor embeddings and the one-hot style label, and output a soft gating vector that blends the experts' outputs.
| Expert Domain | Number of Experts | Expert Names |
|---|---|---|
| Lateral | 5 | sharp left, left, straight, right, sharp right |
| Longitudinal | 3 | braking, cruising, accelerating |
Conditional Diffusion Decoder: The trajectory generation starts from Gaussian noise and iteratively denoises over multiple steps. The diffusion process is conditioned on two things: the fused sensor embeddings from the perception backbone, and the expert-guided latent (the output of the MoE routers). This conditions the denoiser to produce a trajectory that matches the chosen style while respecting the scene geometry.
Style Augmentation: To train the model to understand different styles, the researchers synthetically altered ground-truth trajectories. Aggressive trajectories were created by amplifying speed and extending along the lane centerline, with safety clipping to prevent unrealistic behavior. Conservative trajectories were generated by progressively reducing speed, widening car-following margins, and—if necessary—truncating the horizon to simulate maximum caution.
At inference, the user only needs to pass a style label (0=conservative, 1=neutral, 2=aggressive) along with the sensor data. The routers automatically activate the appropriate experts and guide the diffusion process accordingly.
Why This Matters for Robotics
Style-controllable autonomous driving has direct commercial and operational value. For warehouse robots and logistics vehicles, the ability to switch between cautious and assertive driving modes means the same robot fleet can navigate tight aisles during peak hours and cruise efficiently during off-peak. In passenger vehicles, OEMs can offer a single planning stack that adapts to driver preference or regional driving culture.
The dual MoE architecture is also a powerful template for other robotics domains. Any system that must balance multiple behavioral objectives—such as a robotic arm balancing speed and precision, or a legged robot adjusting walking gait—can benefit from the disentangled expert approach. The key takeaway: instead of training separate models for each style, you can train one model with style-conditioned experts, enabling on-the-fly behavior switching without retraining.
For maintenance engineers and fleet operators, this means fewer models to deploy, simpler updates, and consistent behavior across diverse environments. The use of a diffusion decoder also produces smoother trajectories than traditional optimization-based planners, reducing wear on actuators and improving efficiency.
Limitations and Open Questions
The current model relies on a single discrete style label. Real-world driving is nuanced—a driver might want aggressive acceleration but conservative lane changes. A continuous or multi-dimensional style embedding would be more flexible.
The synthetic trajectory augmentation is simplistic: aggressive means "speed up and swerve," but real aggressive driving also involves tailgating and late braking. The paper does not show how well the model handles edge-case driving behaviors like emergency evasive maneuvers.
Additionally, the evaluation is limited to the NAVSIM dataset, which contains mostly highway and urban driving in US cities. Performance in dense traffic, non-US road rules, or adverse weather is unknown. The model's real-time inference speed and computational cost on embedded hardware are also not reported.
Frequently Asked Questions
How does the style control work at a high level? You pass a simple integer label (0, 1, or 2) corresponding to conservative, neutral, or aggressive driving. The model's dual routers then select which lateral and longitudinal experts to activate, influencing the trajectory the diffusion decoder produces.
Can the model switch styles mid-scene without retraining? Yes, the style label can be changed at any inference step without any fine-tuning. The model was designed to generalize across styles from the same set of weights.
Does the model maintain safety when set to aggressive mode? Yes. The aggressive mode produces higher speeds and more lane changes, but the model still respects traffic lights, lane boundaries, and collision avoidance. Safety metrics like TTC and collision rates remained near the level of neutral mode.
What data was used to train the style-aware model? The base model was trained on the NAVSIM dataset, which includes hundreds of hours of real-world driving data. To teach styles, the researchers synthetically augmented trajectories—speeding up and re-extending for aggressive, slowing down and widening margins for conservative.
Conclusion
D³-MoE demonstrates that style-controllable autonomous driving is achievable with a single end-to-end model by decoupling lateral and longitudinal decisions into separate expert modules. The approach is practical, building on existing datasets and diffusion-based planning, and opens the door for user-adjustable robot behavior in logistics, last-mile delivery, and personal mobility.