scilwb/JEPE_OOD_explore

JEPE OOD Explore artifact bundle

This dataset repository accompanies https://github.com/lwbscu/JEPE_OOD_explore. It contains portable experiment weights, two derivative off-policy datasets, the frozen Cube expert memory index, final reports, and curated visual evidence.

Project summary

Robust v1 + T2 is the best deployable three-color configuration at 92/92/86% (macro 90.00%). Probe XYZ cost raises paired target-space success to 50/58/52%, but waypoint decomposition falls to 16/12/18%. Official Play training preserves expert behavior (-16.34% stopline; 5.16 mm expert depth-5 error) while leaving planner-candidate error at 85.24/118.04/124.55 mm, so the off-policy training line is a completed negative result. The final diagnosis is in reports/PLANNING_PROBLEM_ANALYSIS.md.

Final OOD and transfer summary

Scope	Result	Interpretation
Robust v1 + T2, Red / Blue-v2 / Yellow-v2	92 / 92 / 86%	Deployable three-color result; macro 90.00%
Probe direct, Red offset-25	94%	Privileged-coordinate diagnostic only
Camera / floor first nonzero tier	54 / 38%	Known visual OOD boundaries
Mass x0.5/x2/x4, Red and Blue-v2	92% throughout	Stable over the measured mass range
Global contact friction	88% minimum	Red x2/x4 loses 4 pp; Blue-v2 remains 92%
Cube-double block-0 zero-shot	4/50 (8%)	Weak composition transfer

The no-augmentation controls (78.00% at 12,732 steps and 77.33% at 16,732 steps) remain well below MaskedAug (87.33%) and Robust v1 (90.00%), excluding simple underfitting as the explanation for the augmentation gains. Full provenance and protocol limitations are in reports/OOD_FINAL_REPORT.md and reports/final_matrix.csv.

Contents

• weights/: ColorAug, MaskedAug, Robust v1, two no-augmentation control checkpoints, off-policy v1 and Play v1 negative-result checkpoints, and the robust XYZ probe. Play v1 is retained for reproduction and is not recommended for deployment.
• datasets/offpolicy_cube_v1/: synthetic-noise off-policy rollouts.
• datasets/offpolicy_cube_v2/: planner-in-the-loop off-policy rollouts.
• memory_index/cube_expert_v1/: expert retrieval index with fixed-50 exclusion metadata.
• reports/: final planning analysis, validation record, terminal reports, six-axis/physical/noise/OOD-detector results, and cube-double transfer.
• evidence/: the same 100 MP4s, comparison sheets, OOD curves, calibration, and machine-readable matrix as GitHub.

Released weights

Path	Size	SHA-256 prefix
weights/coloraug/weights_final.pt	68.9 MiB	acb060577ce83b31...
weights/control_noaugment/weights_final.pt	68.9 MiB	2cfe36941194652d...
weights/control_noaugment/weights_step_12732.pt	68.9 MiB	5f994fc004f8bd11...
weights/maskedaug/weights_final.pt	68.9 MiB	d64501aa8e7dac12...
weights/offpolicy_v1/lr5e6_retry_weights_final.pt	68.9 MiB	ea201b78490c650a...
weights/offpolicy_v1/primary_weights_final.pt	68.9 MiB	4fa2478bddc612cf...
weights/play_v1/weights_final.pt	68.9 MiB	12a7ef9d607b2d98...
weights/probes/robust_v1/robust_v1.pt	0.5 MiB	caa7a92435c01df3...
weights/robust_v1/weights_final.pt	68.9 MiB	cffe41b70ed743c7...

Main paired results

Model / training arm	Red	Blue v2	Yellow v2	Macro
MaskedAug + T2	88%	88%	86%	87.33%
Robust v1 + T2	92%	92%	86%	90.00%
No augmentation, 12,732 steps	86%	74%	74%	78.00%
No augmentation, 16,732 cumulative steps	88%	72%	72%	77.33%

Target tier	Robust latent cost	Probe XYZ cost	Delta
In-box	32%	50%	+18 pp
+5 cm	18%	58%	+40 pp
Fallback support (median 5.57 cm)	12%	52%	+40 pp

The probe table is an intermediate diagnostic, not the final navigation result: the paired 4 cm waypoint chain scores 16/12/18%. See the GitHub reports for exact 50-environment vectors, flip tables, loss curves, model/probe provenance, and the complete causal interpretation.

Provenance and exclusions

The derivative trajectories use OGBench Cube initial states but were newly simulated for this project. Every released training/index artifact records the fixed-50 evaluation exclusion. The original 101.9 GB Cube HDF5, PushT source data, and original Quentinll checkpoint are intentionally not included:

• https://huggingface.co/datasets/quentinll/lewm-cube
• https://huggingface.co/quentinll/lewm-cube
• https://github.com/seohongpark/ogbench

Requested +10/+20 cm target tiers exceeded the real dataset support and are reported only at their common documented fallback (median about 5.57 cm). Absolute paths inside manifests/reports are immutable provenance from the originating machine, not required installation paths.

The official Play source dataset and its 11 GiB local conversion workspace are not redistributed. Only the portable 68.9 MiB final negative-result weight, configuration, code, and audit report are included.

Citation

@misc{jepe_ood_explore_2026,
  title={JEPE OOD Explore: Controlled OOD Studies for a Cube World Model},
  author={LWB SCU},
  year={2026},
  howpublished={https://github.com/lwbscu/JEPE_OOD_explore}
}