CausalDrive splits "where to go" from "how to drive" in end-to-end driving, reaching 91.2 PDMS on NAVSIM v1
Synopsis
End-to-end driving policies usually imitate an expert trajectory as one target, yet similar trajectories can arise from different intentions and execution styles, leaving their contributions entangled. CausalDrive splits supervision in two: a deterministic nominal intention planner trained on style-free counterfactual trajectories built from the navigation command, map topology and ego state, and a diffusion-based style policy that models multimodal residuals between nominal and expert trajectories. It reaches 91.2 PDMS on NAVSIM v1 and 89.5 EPDMS on NAVSIM v2, matching or surpassing the best of the compared methods.
Interpretation
The supervision target in end-to-end driving is split into a nominal intention and a driving style, each supervised separately, instead of one policy learning both from a single expert trajectory. Regression planners predict the complete expert trajectory directly, and generative planners model a multimodal distribution over complete trajectories; neither explicitly separates nominal intention from execution style. On NAVSIM v1, CausalDrive with a ResNet-34 backbone reaches 91.2 PDMS, the same as DiffusionDriveV2 in the table; on NAVSIM v2 it reaches 89.5 EPDMS, above the 87.5 reported for DiffusionDriveV2.
A planner that sees neither surrounding agents nor the expert future supplies the nominal intention, and a diffusion policy adds style as a residual on top of that nominal trajectory. The nominal trajectory is built from the navigation command, map topology and ego state through lane association, topological extension and temporal parameterization, serving as a style-free counterfactual target; the style execution policy conditions on that trajectory and surrounding agents and models the difference between nominal and expert trajectories as a multimodal distribution. In an ablation against a DiffusionDrive baseline, adding only the nominal intention planner raises DAC to 97.3 but drops TTC and NC by 24.0 and 3.6; adding the style execution policy recovers TTC and NC by 25.0 and 3.7, reaching 89.9 PDMS, and the ranker brings it to 91.2.
Perspective
Evaluation runs on the NAVSIM v1 and NAVSIM v2 planning benchmarks, with 1,192 training scenes from navtrain and 136 evaluation scenes from navtest, scored by PDMS and EPDMS; the data-efficiency experiments compare 20%, 40%, 60%, 80% and 100% of expert data, all evaluated on the fixed navtest split.
The data-efficiency advantage narrows as more training data becomes available, and the lead over DiffusionDriveV2 shrinks with abundant data; prioritizing scenes by residual ratio gives a clear gain from the top 40% of scenes, but marginal benefit diminishes as the subset grows and small-residual scenes contribute little.
