Repository navigation
10 bernhard dev Cosim PR - #597
Merged
Merged
Conversation
Also fix bug where lane assignment didn't have a z filter.
eval.reward_lane_center=0.0075 \
configurable and change defaults for the next eval.
… visualization script.
…dropout. Fix policy not seeing far ahead on map.
…rDrive into 10_bernhard_dev
eugenevinitsky
approved these changes
Oct 5, 2026
eugenevinitsky
left a comment
eugenevinitsky
left a comment
There was a problem hiding this comment.
Approved, if possible please tag with a wandb run for posterity
Collaborator
Author
|
Already added the wandb run above. thought the link is now coreweave (think wandb got bought) |
Closed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This PR contains everything for the cosim of nuPlan and CARLA as well as related policy changes to push the performance of the policy on these benchmarks.
In particular:
Speed limit randomization via speed zones.
The CARLA maps have a limited range of speed signs (30km/h, 108Km/h) which doesn't cover the whole distribution and leads to issues on nuPlan and Alpasim. I added speed zone which randomize speed limits plus minus 30% to get better coverage.
Stop signs
Added stop sign to ego input and reward. This is needed for CARLA.
SPS optimizations. Improved SPS by 25% by using pinned memory
Add nuPlan cosim code
Add CARLA cosim code
Some smaller bug fixes
Remove 20% of traffic light groups in training like in GIGAFLOW paper
Optimize Map cache to use less RAM
Also includes Valentins truncation changes
Add k_scaled_0036 weights our best model in self-play so far.
Add speed limits to the lane observation to enable the agent to break preemptively for speed limit changes, avoiding penalty.
This PR includes all the changes + improvements of #562 which should be closed after this is merged.
Latest Run:
https://forge.coreweave.com/wandb/emerge_/nightly-multi-long/runs/k_scaled_0044_1000?nw=nwuserbern_jaeger
Current evals:
Self-play: 718k km/infraction (similar to before)
nuPlan reactive: 91.8 CLS (up from 90 CLS)
nuPlan nonreactive: 88.2 CLS (new)
longest6v2: 48 DS (up from 35 DS)
Ready to merge now