Skip to main content
Back to timeline
arXivSource publication:

A world model can predict the future and still drive out of bounds: PGSD distills search repair into one verification rollout, gaining 11.24 EPDMS on NAVSIM v2 navhard

Synopsis

World-model planners can predict future scenes yet cannot guarantee that the chosen trajectory stays inside road boundaries once a tracking controller executes it, and re-ranking cannot help when no candidate in the pool is compliant. PGSD adds execution-aware selection on top of a frozen WorldDrive planner and FAR rewarder, then distills a search teacher that picks targets by executed road clearance into a future-conditioned repair student that runs one verification rollout per trigger. Single-view PGSD reaches 90.97 PDMS on NAVSIM v1 and 38.96 EPDMS on NAVSIM v2 navhard, 11.24 points (40.5%) above WorldDrive, retaining 73.2% of the search teacher's gain over execution-aware selection.

Source-provided article image: Efficient World-Model Planning via Physics-Guided Search Distillation
arXiv

Interpretation

The planner no longer scores candidate trajectories only by predicted futures; it checks in simulated execution whether the ego footprint crosses a road boundary and prefers the candidate that can actually be driven. Earlier world-model planners re-ranked candidates with future-aware scoring, but ordering cannot create a compliant trajectory when the candidate pool contains none. A diagnostic over all 12,146 NAVSIM v1 scenarios found 278 requiring intervention, 106 of them (38.1%) with no road-compliant in-pool alternative; adding execution-aware selection raised PDMS from 89.0 to 90.4.

For triggered cases with no in-pool solution, the student predicts one beyond-pool lateral correction instead of searching online repeatedly. Physical search finds such corrections but costs repeated rollouts per trigger, while imitating search directly lets teacher trajectories that barely clear the constraints amplify prediction error. In a paired comparison of 106 repair triggers, search used 1,040 additional rollouts versus 106 for the student, an 89.8% reduction; FCRS reaches 90.97 PDMS, within 0.21 points of the search teacher CAST at 91.2, with median cached repair latency of 25.13 ms versus 239.49 ms.

Teacher demonstrations are selected by minimum executed road clearance, giving the student a geometric buffer so imitation error does not push the correction back across the boundary. Covering the same 680 of 773 infeasible training references and changing only target selection raised median target clearance from 9.8 to 31.3 cm. On 241 development references, repair success rose by 17.15 percentage points and rejection fell from 47.0% to 29.9%; on the same 202 teacher-feasible references, success rose from 62.7% to 82.7%.

Perspective

The demonstrated setting is single-view visual input with map-based execution geometry, evaluated on all 12,146 NAVSIM v1 navtest scenarios and on NAVSIM v2 navhard_two_stage (5,912 frames, 225 groups), with execution simulated by an LQR tracker and kinematic bicycle model; HUGSIM Hard supplies only qualitative zero-shot examples.

The authors list richer maneuvers and interaction-aware constraints as future work; negative-clearance cases remain even after clearance supervision improves, so verification and fallback are kept, and fallback is not counted as successful repair. The controlled removal of execution feedback supports it as a group of inputs without isolating individual channels.