Anticipate the truck before it turns

Model output at this observation

Held-out length · fixed final checkpoint

Training uses 16.1 m and 20.1 m trucks. All 18.1 m episodes, appearances, stopping branches and temporal frames are reserved for testing.

What this experiment measures

48 unique observations (32 train, 16 test) across 18 episodes. Each input contains three current cameras and four numeric history states at 2 Hz; the model predicts four future XY points at 1 Hz. Before braking, the three possible stopping branches share one observation and are retained as admissible alternative futures.

All three reference stops are geometrically safe for every tested truck length. Reference trajectory sets are shared across training and test lengths. The experiment tests anticipation and length interpolation within this scene; it does not establish a unique optimal stop or generalization to another road. The source is a NAVSIM test scene, reused for this custom split, so these results are not official NAVSIM benchmark scores.

Truck dimensions, hitch motion, road contact and reference clearances are checked in 3D. Generated photographs approximate camera projection and can alter some background pixels. References use idealized constant braking; they are not jerk-validated. Four-second open-loop clearance does not establish safety through the entire later turn.