Skip to content

Tmp #301

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
daphne-cornelisse wants to merge 128 commits into
base: 2.0
Choose a base branch
from
Draft

Tmp #301

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
128 commits
Select commit Hold shift + click to select a range
efd616d
Implement customizable action space, truncation support and new goal …
daphne-cornelisse Feb 11, 2026
ec3f22f
Implement truncation logic in train loop and improved realism evals.
daphne-cornelisse Feb 12, 2026
944f8e1
By default, use term mode 1 for train (truncated episodes) and 0 for …
daphne-cornelisse Feb 12, 2026
a139fd2
Implement customizable action space, truncation support and new goal …
daphne-cornelisse Feb 11, 2026
dd0032c
Implement truncation logic in train loop and improved realism evals.
daphne-cornelisse Feb 12, 2026
f7174d3
By default, use term mode 1 for train (truncated episodes) and 0 for …
daphne-cornelisse Feb 12, 2026
0790ed0
Merge branch 'gsp_v0' of https://github.com/Emerge-Lab/PufferDrive in…
daphne-cornelisse Feb 13, 2026
7b726e5
Make agents type-aware and ignore pedestrian off-road events/collisio…
daphne-cornelisse Feb 13, 2026
4918229
Human regularization beta version - still some instabilities during t…
daphne-cornelisse Feb 14, 2026
0dd95a9
Merge remote-tracking branch 'origin/2.0' into gsp_v0
daphne-cornelisse Feb 16, 2026
f537f95
Bug fix: use a lightweight c_step function when collecting expert dat…
daphne-cornelisse Feb 16, 2026
6045c2f
Minor improvements and robustness.
daphne-cornelisse Feb 16, 2026
d9d4335
Improve regularization logging.
daphne-cornelisse Feb 16, 2026
cfaba67
Improve the bins -> more accurate inference of actions.
daphne-cornelisse Feb 16, 2026
bc5d2e4
Fix WOSAC (filtering out dones currently unsupported).
daphne-cornelisse Feb 16, 2026
7e4d292
Allow for controlled experiments over non-numeric values
daphne-cornelisse Feb 16, 2026
caf86ee
Support for human regularization with memory/LSTM. Also a significant…
daphne-cornelisse Feb 16, 2026
ac2c565
Fix stupid rendering bug.
daphne-cornelisse Feb 20, 2026
a6add94
Merge branch '2.0' of https://github.com/Emerge-Lab/PufferDrive into …
daphne-cornelisse Feb 20, 2026
3dc1cdd
Setting improvements.
daphne-cornelisse Feb 20, 2026
fa9bea1
Fix mismatch between renderer and env logic.
daphne-cornelisse Feb 21, 2026
36c2755
Comment out LSTM
daphne-cornelisse Feb 21, 2026
f4c8e92
Make control mode and agents configurable for rendering.
daphne-cornelisse Feb 22, 2026
0124fd6
Bug fix for stop goal behavior.
daphne-cornelisse Feb 22, 2026
504beeb
WIP
daphne-cornelisse Feb 25, 2026
5c2366b
Merge remote-tracking branch 'origin/2.0' into gsp_v0
daphne-cornelisse Feb 26, 2026
cc501a1
IL training
daphne-cornelisse Feb 26, 2026
0be23aa
Reg signs of life.
daphne-cornelisse Feb 27, 2026
6a0fcc2
Implement KL anchor: Guidance is KL div to pretrained BC policy.
daphne-cornelisse Feb 27, 2026
7bcf827
Implement BC training with resampling.
daphne-cornelisse Feb 27, 2026
0821892
Improvements in anchor settings.
daphne-cornelisse Feb 27, 2026
3b0b202
Filtering
daphne-cornelisse Feb 28, 2026
d189994
Merge branch 'filter_transitions' into gsp_v0
daphne-cornelisse Feb 28, 2026
88b836d
Write script to filter for interactive SDC scenes. Need better human-…
daphne-cornelisse Mar 1, 2026
d600762
Focus camera on SDC in human-replay model.
daphne-cornelisse Mar 1, 2026
7c8b3aa
Script to filter for interactive scenarios. Also update BD policy: fo…
daphne-cornelisse Mar 1, 2026
7a0d44d
Save good configs: observe clear effect with conditioning.
daphne-cornelisse Mar 1, 2026
7f97285
Implement lambda conditioning: Create population of agents during tra…
daphne-cornelisse Mar 2, 2026
a172484
Fix human-replay rendering: Stop when SDC reaches destination.
daphne-cornelisse Mar 2, 2026
0224cd1
Ensure correct goal behavior for human-replay eval.
daphne-cornelisse Mar 2, 2026
4aa3f2c
Make the arrow head more pronounced and reflect speed.
daphne-cornelisse Mar 2, 2026
8f9b975
Add eval to check the effect of different lambda values: Looks like i…
daphne-cornelisse Mar 2, 2026
e56eebb
log
daphne-cornelisse Mar 4, 2026
c370fdb
Merge remote-tracking branch 'origin/2.0' into gsp_v0
daphne-cornelisse Mar 4, 2026
dcc68fe
Fix memory leak -> free expert data arrays.
daphne-cornelisse Mar 4, 2026
436c877
Lambda sweep only occassionally.
daphne-cornelisse Mar 4, 2026
df2493e
Improved version of the lambda conditioning to ensure within env dive…
daphne-cornelisse Mar 4, 2026
7510f98
Implement reward conditioning.
daphne-cornelisse Mar 5, 2026
5fdaf6e
Replace hr_collision_rate with valid_collision_rate: only count colli…
daphne-cornelisse Mar 6, 2026
99d48f6
Feature: Add expert_replay and inferred_expert_actions control modes …
daphne-cornelisse Mar 7, 2026
9e29068
Implement delta dynamics model.
daphne-cornelisse Mar 7, 2026
ec68aa0
Small fixes to enable kl-regularized training with delta model.
daphne-cornelisse Mar 7, 2026
407fbfd
Several small improvements.
daphne-cornelisse Mar 9, 2026
63b861a
Small improvements.
Mar 10, 2026
148a84a
Delete old models.
daphne-cornelisse Mar 10, 2026
b3a0133
Teleport agents each steps for data collection.
daphne-cornelisse Mar 11, 2026
daab7c3
Fix the lambda value.
daphne-cornelisse Mar 11, 2026
9e04716
Change conditioning ranges.
daphne-cornelisse Mar 11, 2026
edb0441
Log the unweighted kl div.
daphne-cornelisse Mar 11, 2026
8c46cc9
Small changes
daphne-cornelisse Mar 13, 2026
abba153
Add anchor comment for clarity
daphne-cornelisse Mar 13, 2026
884a5bc
Merge remote-tracking branch 'origin/2.0' into gsp_v0
daphne-cornelisse Mar 13, 2026
5d4e698
Fix wheelbase for classic and draw ADE so we can filter out data.
daphne-cornelisse Mar 14, 2026
b53ae7e
Include data filtering rules so that we don't train the BC policy on …
daphne-cornelisse Mar 14, 2026
efc25ac
Set ADE threshold to 3
daphne-cornelisse Mar 14, 2026
4d762a9
Fix wheelbase.
daphne-cornelisse Mar 15, 2026
1469d4d
Perfect BC policy on a single map.
daphne-cornelisse Mar 16, 2026
df63e8f
Scaling up to 100 scenes and more detailed eval stats.
daphne-cornelisse Mar 16, 2026
ff287c9
Increase lambda-conditioning ranges.
daphne-cornelisse Mar 16, 2026
4a0d7e6
Apply Gigaflow-style sensor noise.
daphne-cornelisse Mar 16, 2026
0d5ed3f
Add option to visualize sensor noise + turn off for now.
daphne-cornelisse Mar 17, 2026
5f88ac5
Major env cleanup. Getting rid of respawning logic and dead code. Val…
daphne-cornelisse Mar 18, 2026
a5f342c
Updated configs
daphne-cornelisse Mar 18, 2026
31465df
Re-implement delta-dynamics model for BC.
daphne-cornelisse Mar 18, 2026
0dd593f
Minor
daphne-cornelisse Mar 18, 2026
68e122e
Merge branch 'gsp_v0' of https://github.com/Emerge-Lab/PufferDrive in…
daphne-cornelisse Mar 18, 2026
93fd6e1
Bug fix for delta: Use a multi-discrete independent action space.
daphne-cornelisse Mar 18, 2026
372b3a7
Upgrade BC policy training script: now supporting sweeps. Best perf w…
daphne-cornelisse Mar 19, 2026
7b7ff4f
Improve BC policy.
daphne-cornelisse Mar 19, 2026
253e7e9
Log entropy of individual action heads as sanity check.
daphne-cornelisse Mar 19, 2026
1d41a95
Add validation loop for BC training.
daphne-cornelisse Mar 19, 2026
f3f6277
Focus self-play rendering on SDC and updated settings.
daphne-cornelisse Mar 19, 2026
09c1236
Some minor emaining render bug fixes and make sensor noise configurab…
daphne-cornelisse Mar 19, 2026
a0a35aa
Update settings.
daphne-cornelisse Mar 19, 2026
5ebf48c
Support sweeps with any specified metrics. Tested with eval/hr_mean_d…
daphne-cornelisse Mar 20, 2026
6ab6994
Fix defaults.
daphne-cornelisse Mar 20, 2026
dce7f90
Fix bug in sweep code.
daphne-cornelisse Mar 20, 2026
f7c222d
Temp
Mar 20, 2026
09a947e
Bug fix in sensor noise config.
daphne-cornelisse Mar 21, 2026
bcb064a
Couple of minor fixes.
daphne-cornelisse Mar 21, 2026
7c918c6
Merge branch 'gsp_v0' of https://github.com/Emerge-Lab/PufferDrive in…
daphne-cornelisse Mar 21, 2026
5b4aecd
Minor
daphne-cornelisse Mar 21, 2026
ee9531c
New hparam ranges.
daphne-cornelisse Mar 21, 2026
55d30a3
Configs.
daphne-cornelisse Mar 21, 2026
30d5654
Enable WOSAC eval.
daphne-cornelisse Mar 24, 2026
ada2adc
Fix bug in WOSAC logging.
daphne-cornelisse Mar 24, 2026
da53999
Bug fix: Make sure to update the expert replay velocities.
daphne-cornelisse Mar 26, 2026
7b40225
Data updates and quality improvement.
daphne-cornelisse Mar 26, 2026
32521ad
Minor; bc improvments; 91% acc, decent cl perf.
daphne-cornelisse Mar 26, 2026
92bfa30
New defaults, getting good agents 1% hr coll and vibe check.
daphne-cornelisse Mar 27, 2026
b95a49a
initial commit (#367)
julianh65 Mar 27, 2026
aa75f46
Analysis pipeline v1.
daphne-cornelisse Mar 27, 2026
6d6780e
Analysis pipeline v2.
daphne-cornelisse Mar 27, 2026
ee52aaf
Improvements to filtering script.
daphne-cornelisse Mar 27, 2026
a2fa1fc
Analysis pipeline v2 - with figures.
daphne-cornelisse Mar 27, 2026
ae42df0
Set goal radius to 2.5 by defaul.t
daphne-cornelisse Mar 27, 2026
139c554
BC policy update.
Mar 28, 2026
d54c107
Make map sampling deterministic.
daphne-cornelisse Mar 29, 2026
9bf9405
Automatically select folder with maps based on num_maps.
daphne-cornelisse Mar 29, 2026
197657e
Minor
daphne-cornelisse Mar 29, 2026
ac8545a
Fix rendering.
daphne-cornelisse Mar 29, 2026
4c3d242
New BC anchor checkpoints.
daphne-cornelisse Mar 29, 2026
19c8ef8
render fixes.
daphne-cornelisse Mar 30, 2026
51f551c
Temp fix for eval.
daphne-cornelisse Mar 30, 2026
5e4adfc
Bug fix: Make sure to actually increment the resample count.
daphne-cornelisse Mar 30, 2026
1ae4fcf
Upload cpt to wandb.
daphne-cornelisse Mar 30, 2026
68b271d
Add metadata scaling analysis.
daphne-cornelisse Mar 30, 2026
330f5b4
Update anchor policies
Mar 31, 2026
1dcb497
Placeholder bc classic pol.
daphne-cornelisse Mar 31, 2026
16d4031
Add in scaling analysis and wosac eval.
daphne-cornelisse Mar 31, 2026
2e05988
Separate ffmpeg recorder from raylib window lifecycle to support rend…
daphne-cornelisse Apr 1, 2026
3879b90
Minor updates.
daphne-cornelisse Apr 1, 2026
ef38a2d
Add support for async_reset arg. False is used for evals.
daphne-cornelisse Apr 2, 2026
028c6e4
Classic runs
daphne-cornelisse Apr 2, 2026
943a66e
Eval bug fix: Zero out the logs after aggregating. This bug affects a…
daphne-cornelisse Apr 2, 2026
c06a430
Eval improvements: Obtain 0.0% at fault coll rate with best reg policy.
daphne-cornelisse Apr 2, 2026
c92c40e
Add classic dynamics model to eval.
daphne-cornelisse Apr 2, 2026
02ddf77
Classic baseline configs.
daphne-cornelisse Apr 3, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions .gitignore
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -190,3 +190,7 @@ pufferlib/resources/drive/output*.gif
emsdk/
docs/book/*
!docs/book/assets/

logs/*

eval_videos/*
11 changes: 11 additions & 0 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -216,3 +216,14 @@ If you use PufferDrive in your research, please cite:
year = {2025},
}
```


## Features

### Verify human data

Run
```
puffer verify puffer_drive --env.render-mode 0
```
with `control_mode` set to `expert_replay` or `inferred_`. Note: Currently only supported with classic dynamics model.
Loading
Loading