OpenPI pi0.5 oven subtask full100 overhead 20k artifacts
Artifacts for the corrected-subtask OpenPI subtask run:
- Run ID:
vladaptor_oven_pi05_keypose_subtask_refined_arm_full100_overhead_b16_20k
- Config:
pi05_vladaptor_oven_keypose_subtask_full
- Base:
pi0.5
- Dataset:
lsnu/vladaptor_take_tray_out_of_oven_keypose_full100_overhead
- Training split: full 100 demonstration episodes
- Global camera:
overhead
- Checkpoints: 1k, 5k, 10k, 15k, 20k
- Repo created for archival/reproducibility, not for polished release.
Directory layout
checkpoints/step_01000: 1k checkpointcheckpoints/step_05000: 5k checkpointcheckpoints/step_10000: 10k checkpointcheckpoints/step_15000: 15k checkpointcheckpoints/step_20000: 20k checkpointdataset/: LeRobot dataset, overhead renders, and corrected subtask labelsdataset/raw_vladaptorbench_lowdim_bimanual_take_tray_out_of_oven_train_128_all_variations: raw low-dim source episodes used for dataset generationlogs/: training logs, sim eval logs, summaries, policy server logs, subtask legibility checkscode/: OpenPI subtask source snapshot, scripts, and eval Python scriptscode/uv.lock: OpenPI lockfile from the training checkoutcode/sim_runtime_repos_metadata: RLBench/PyRep/YARR/peract repo heads and dependency fileseval_artifacts/: seed eval sets and calibration artifacts used during sim evaldiagnostics/dataset_preparation: dataset/camera smoke images and preparation summariesmetadata/: file manifest, git status/diff, run summary, subtask list, environment snapshot, base parameter manifest, and provenance notesprior_open_more_v17/: older Open More / reach-and-regrasp v17 context, code, audits, previous-version metadata, and a reconstruction_bundle/ with the old checkpoints/data/code needed to rerun the v17 wrapper stack. This prior line used a different front_rgb Open More dataset with 80 training episodes, not the current overhead/full100 dataset.estimator_oracle_adaptor_20260428/: later estimator/oracle adaptor analysis bundle and raw artifacts, including DINOv3 estimator checkpoints/features, Kaggle DINOv3 Keras backbone weights, oracle adaptor ablations, and neutralized estimator-adaptor analysis notes.
Current Context Addendum - 2026-04-28
This repository now contains multiple dated experiment lines. The original
sections below describe the base pi0.5 subtask checkpoint training/evaluation
artifact. Later adaptor analysis is under
estimator_oracle_adaptor_20260428/.
Historical documents may use words such as current, best, or final inside
their dated experiment context. Those labels are not repo-wide claims after the
2026-04-28 estimator/oracle adaptor upload.
For the latest neutralized estimator-adaptor context, start with:
estimator_oracle_adaptor_20260428/analysis_bundle/estimator_oracle_adaptor_analysis_bundle_20260428.zipestimator_oracle_adaptor_20260428/non_oracle_estimator_adaptor/docs/neutral_analysis_context_20260428.mdestimator_oracle_adaptor_20260428/non_oracle_estimator_adaptor/docs/dinov3_adaptor_eval.mdestimator_oracle_adaptor_20260428/non_oracle_estimator_adaptor/docs/contact_valid_repeatability_20260428.md
The 15k checkpoint statement below is scoped to the base checkpoint eval table
in this README. It does not include later oracle or estimator adaptor results.
Corrected subtask list
10 approach the oven and align the right gripper with the handle
21 use the right arm to open the door using the handle
32 use the right arm to open the door using the handle
43 use the right arm to pull the door open and release the handle
54 use the right arm to pull the door open and release the handle
65 move the right arm clear of the oven opening
76 move the left gripper to the tray and grasp it
87 move the left gripper to the tray and grasp it
98 use the left arm to pull the tray out of the oven
109 use the left arm to lift the tray up out of the oven
1110 use the left arm to release the tray and finish
1211 use the left arm to release the tray and finish
1312 use the left arm to release the tray and finish
1413 use the left arm to release the tray and finish
Eval protocol
Checkpoint evals used live RLBench seed episodes:
- Seeds:
100000 through 100019
- Episodes: 20
- Variation index: 0
- Global camera:
overhead
actions_per_infer: 1- Max env actions: 80
- Collision mode: ignore collisions
Sim eval results
| Step | Success | Tray touch | Errors | Mean env actions |
|---|
| 1k | 0/20 (0%) | 4/20 (20%) | 0 | 80.00 |
| 5k | 3/20 (15%) | 11/20 (55%) | 0 | 77.25 |
| 10k | 7/20 (35%) | 11/20 (55%) | 0 | 58.65 |
| 15k | 9/20 (45%) | 12/20 (60%) | 0 | 53.70 |
| 20k | 3/20 (15%) | 12/20 (60%) | 0 | 70.80 |
Within the base checkpoint eval table above, the 15k checkpoint had the highest
task-success count. The 20k checkpoint had lower task success than 15k while
tray-touch rate stayed flat.
Subtask legibility checks were run at all saved checkpoints. Generated subtask logs are included under logs/; some generated strings include formatting suffixes such as .; Action:.
Code provenance
The OpenPI subtask source snapshot and local scripts are included under code/. Git head, remote, status, and the worktree diff at packaging time are in metadata/.
The pi0.5 base parameter URL/size manifest used to hydrate the run is included at metadata/pi05_base_params_http_manifest.tsv. The final checkpoints are full checkpoints and do not require separately reconstructing the base checkpoint for inference or resume.
Older Open More / regrasp context is separated under prior_open_more_v17/ because it was trained and evaluated on a different dataset/camera/split from this run.
No license is specified here.
