Investigate map-input sensitivity in trajectory planner (split from #168)

[ZaynabEM]

Description

Split from #168 (Federated Model Training); logged separately since the map/BEV investigation has outgrown the training-benchmark issue it started in.

Original #168 scope already flagged the map choice: "Consider map representation, rendering etc. we want to use navigation maps not Lanelet2 maps for now."

Findings logged against #168 (WG meeting, 19 Aug 2026, Slot 1 — opened explicitly "to solve issue of bottleneck for the BEV part"):

• For this experiment, the map input was simplified: instead of SDTagNet, a black background plus route information ("tell model where to drive") was used
• Map backbone tested frozen or trainable, on a black-and-white pretrained ResNet50
• Noted that using pre-trained frozen BEV modules may require changing the BEV size

Findings logged against #168 (WG meeting, 26 Aug 2026, Slot 1):

• Stated objective: the model should build a local map from the navigation map, not rely on a map a priori
• SDTagNet takes the navigation map as an additional prior input
• Gap identified: route definition is missing in the current trajectory-planner architecture

Related finding, same meeting (26 Aug 2026, Slot 2) — flagging for confirmation, since this row isn't explicitly tagged to #168 in the tracker the way Slot 1 is:

• Zeroing the map input significantly impacts model performance, while zeroing camera input has minimal effect
• Map input currently dominates BEV formation, overshadowing perception contributions

More context:
The tracker shows a longer evolution of map handling before this issue existed: the raw Lanelet2 HD map was dropped for a rendered 2D path / 4-channel RGB-route tensor (22 Jul), a dedicated map encoder was added under Model/model_components/map_encoder (#47, 17 Jun), map features were found to possibly confuse the model — performance was slightly better without map input (05 Aug) — and the route encoder was later split out from the map encoder (12 Aug).

Questions to resolve:

• Is the strong map-dependence found on 26 Aug consistent with, or in tension with, the "map features may confuse model" finding from 05 Aug?
• Route-definition gap in the trajectory-planner architecture (26 Aug) need to be fixed before the map-sensitivity question can be answered cleanly?
• Does this hold across both trajectory planner variants in Federated Model Training #168 (Bezier, Flow Matching)?

@m-zain-khawaja @riita10069 @gcordova10 @TranHuuNhatHuy