Two trailers · 28.5 metres

Model output at this observation

Frozen models · no retraining

Baseline, Stage 1 and DriveMA-B are the existing models from the corrected single-trailer experiment. All eight observations here are test-only. Their weights are pinned before image generation.

3D rehearsal · orbit and play

What this test measures

Eight unique observations, three stopping branches, and 24 current camera images. Each input contains three current cameras plus four numeric history states at 2 Hz. The output is four future XY points at 1 Hz. At t ≤ 0, the three ego stop choices share one observation and are scored as admissible alternatives.

The 28.5 m combination has two independently articulated trailers. Both are included in collision checks. Minimum modeled reference clearance is 0.405 m, 0.278 m and 0.136 m for stops at 10.0 m, 11.3 m and 12.6 m. These tight margins are not a real-world safety guarantee.

This is a within-scene extrapolation test on authored geometry and generated photographs, not an official NAVSIM score or a new-road test. Photographic projection is approximate. Reference braking is idealized; four-second open-loop clearance does not establish later or closed-loop safety.