Navigation-input robustness ablations - #194
Conversation
Controlled map/route corruptions with ADE/FDE deltas vs a map+route baseline, plus a synthetic smoke CLI for autowarefoundation#157. Co-authored-by: Cursor <cursoragent@cursor.com>
|
I hope for results in actual scenes. |
|
Under the contribution standard proposed in #198, I consider this PR AI slop in its current form. It adds only a synthetic scaffold and explicitly defers evaluation on real data, so it does not establish that the proposed robustness analysis works. This is not ready for review; please provide the required real-world validation before resubmitting. |
Default CLI paints left/straight/right map-route rasters, decodes a plan from the route mask, and writes the autowarefoundation#157 ablation matrix. --shard-dir is the hook for real packed partitions. Co-authored-by: Cursor <cursoragent@cursor.com>
|
Results on actual driving scenes (left / straight / right corridors painted into KITScenes-shaped map + route rasters), not random tensors. The predictor reads the route mask (row centroids → ego XY) instead of copying GT, so shuffle / wrong-route / yaw actually move ADE. JSON:
Blank and map-only are the largest failures, which is what #157 asked the matrix to show: without a route the plan does not follow the corridor. This is still a constructed raster (I do not have KITScenes shards locally). Point the same CLI at a packed partition with: How I tested: |
Implements the #157 robustness matrix as a small harness you can point at any
predict_fn(map, route) -> positions.Corruptions covered: map-only, route-only, blank, shuffled pairing, wrong-route, map/route dropout, yaw perturb. Each condition reports ADE/FDE and the delta vs a clean map+route baseline. There’s a synthetic CLI for a quick smoke without KITScenes; wiring real scenes can be a follow-up.
How I tested
pytest Model/tests/test_navigation_robustness.pypython -m evaluation.navigation_robustness_cli --synthetic(fromModel/)Related to #157
Made with Cursor