Skip to content

Bugfix: Lateral Corridor Exit - #215

Merged
mwatson-nvidia merged 1 commit into
e2e_challengefrom
e2e_challenge_fix_lateral_corridor_eval
Oct 20, 2026
Merged

mwatson-nvidia merged 1 commit into
e2e_challengefrom
e2e_challenge_fix_lateral_corridor_eval

Conversation

@mwatson-nvidia

Copy link
Copy Markdown
Collaborator

As described in #198 , the direction calculation used to determine whether a vehicle exited the corridor laterally was not robust to cases where the vehicle stopped at the end of the trajectory. This change addresses that gap by ensuring at least 1 meter exists between trajectory points when computing this direction vector.

@copy-pr-bot

copy-pr-bot Bot commented Oct 20, 2026

Copy link
Copy Markdown

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

@mwatson-nvidia
mwatson-nvidia merged commit 1b024f0 into e2e_challenge Oct 20, 2026
1 check passed
@mwatson-nvidia
mwatson-nvidia deleted the e2e_challenge_fix_lateral_corridor_eval branch October 20, 2026 10:16
rdesc added a commit to rdesc/alpasim that referenced this pull request Oct 20, 2026
Brings in NVlabs#215 (Bugfix: Lateral Corridor Exit) and NVlabs#217 (Enable report
uploading).

NVlabs#215 matters for our completed runs. left_corridor_laterally is a binary
pass/fail criterion and accounted for 64% of failures in the curated_val
run, so scenes where the recording ends with the ego stopped may have been
misclassified: the old _end_direction() used only the final two GT
coordinates, which are duplicates or jitter for a stopped vehicle, making
longitudinal overshoot read as lateral departure.

Only a scorer changed, so completed runs can be rescored from their ASLs
with alpasim-reeval; no re-simulation needed.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
rdesc added a commit to rdesc/alpasim that referenced this pull request Oct 20, 2026
Upstream NVlabs#215 (1b024f0) landed after our run. left_corridor_laterally is a
binary pass/fail criterion, and the old direction calculation used the last two
ground-truth coordinates, which are duplicates or jitter when the recording
ends with the ego stopped -- scoring longitudinal overshoot as a lateral
corridor exit.

Rescored from the existing ASLs with alpasim-reeval (scorer-only change, no
re-simulation): 250 -> 257 scenes passed, 122 -> 115 corridor failures, mean
scene score 0.5455 -> 0.5614. Seven scenes were false positives. The earlier
finding holds: corridor exit alone still accounts for 26% of scenes and 60% of
failures.

The Drive-IRT ranking deliberately stays on pre-fix scores. The bundled
reference summaries predate NVlabs#215 and cannot be rescored (we have their
results-summary.json, not their ASLs), so mixing post-fix ours with pre-fix
theirs would credit us with false positives removed on one side only.

Also records agreement with OPTED (arXiv:2609.20756), which evaluates Alpamayo
1.5 on the same 441 scenes: every closed-loop metric lands within or at their
error bars. alpamayo15_curated_val_table.tex is a LaTeX table in their Table I
format, with ADE and latency left blank since we measured neither comparably.

fix_run_logs.sh repairs a shared-log problem: launch_train.sh appended two runs
to one train_run.log, so per-scene counts scraped from it mixed the aborted
sample400 run with mix200. It splits the log per run and patches the launcher
to log into each run directory.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant