Οβ.β training ablations on UMI tossing (PLAN D21)
Four step-matched Οβ.β
full-SFT runs (10,000 steps Γ batch 64 = 640,000 samples, a third of the main run) that vary the training data or sampling. Each folder is an openpi inference bundle like the main checkpoint (params/, assets/, export_manifest.json), exported at the run's selected checkpoint. Lower composite is better (validation fast preset, 7,166 anchors).
| Folder | Change | Selected step | Norm stats (assets/) |
Val composite: all / dynamic / tennis |
|---|---|---|---|---|
pi05_abl10k_ref_v1_best/ |
reference: whole train split, uniform sampling | 9999 | umi_toss_v1 |
0.2199 / 0.2137 / 0.2628 |
pi05_abl10k_release_v1_best/ |
release-centred sampling | 9999 | umi_toss_v1 |
0.2149 / 0.2091 / 0.2583 |
pi05_abl10k_dyn_v1_best/ |
dynamic-tossing episodes only | 9999 | umi_toss_v1_dyn |
0.3532 / 0.2084 / 1.3215 |
pi05_abl10k_ten_v1_best/ |
tennis episodes only | 9999 | umi_toss_v1_tennis |
0.6621 / 0.7229 / 0.2581 |
Each arm is a single run (one seed) with no confidence interval, so none of these differences is established as an effect (docs/REPORT.md Β§7.1).
How to load
Load each folder like the main Οβ.β
checkpoint (policies/pi05/adapter.py, openpi 215abfb + the benchmark's patches), with the main config --config pi05_umi_tossing (the adapter default). The ablation configs share its model and transforms, and the adapter takes the norm stats from the bundle's own assets/<asset_id>. The config names in export_manifest.json (pi05_abl10k_*) are registered only by training.ablations.pi05_configs.register(); passed as --config to the adapter they fail with Config ... not found.
Licence
Gemma model derivatives: Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms. A copy is included as LICENSE_GEMMA.txt.
Provenance
- Code, configs, records and the full report:
nrvna-ai/umi-tossing-benchmark(private repository) at commit454f50306e2f. - The scores here are offline, open-loop validation scores against held-out recorded human demonstrations (val
fast, 7,166 anchors). These checkpoints were not run in simulation or on a real robot. - Each bundle's params and norm stats are byte-identical to the scored training checkpoint; these files are hard links of the exports.
Files
94 model files, 49.77 GB in total. SHA256SUMS lists the SHA-256 of each of them (verify with sha256sum -c SHA256SUMS; it does not cover this README or the licence file). After upload, every file was checked against the Hub: LFS files by SHA-256, small files by git blob hash.
| File | Size | SHA-256 |
|---|---|---|
pi05_abl10k_ref_v1_best/params/ocdbt.process_0/d/eba36210edd5932d9189b34dea93fc2c |
3.22 GB | 96c4e8a2714e4126β¦ |
pi05_abl10k_dyn_v1_best/params/ocdbt.process_0/d/9aa098a57ed34910e904d6334702316d |
3.01 GB | 791d1f47823413f7β¦ |
pi05_abl10k_release_v1_best/params/ocdbt.process_0/d/3ee9ab9ed4cbbb1d2a1e7bbed37b3427 |
3.01 GB | a40ca31128adc8b8β¦ |
pi05_abl10k_ten_v1_best/params/ocdbt.process_0/d/6ce82508fd2c6c4ae9a95f1896a3a492 |
2.76 GB | 4d7cf4dd790a5640β¦ |
pi05_abl10k_ref_v1_best/params/ocdbt.process_0/d/4d85a44036970f8c08d28c52aa89b18d |
2.53 GB | bf17164662aa4164β¦ |
pi05_abl10k_ten_v1_best/params/ocdbt.process_0/d/977d8fee428fb2d1294d1c3cb3c3e96d |
2.39 GB | 984c012229a69d88β¦ |
pi05_abl10k_ten_v1_best/params/ocdbt.process_0/d/d51b77826a1414307d5148dcedd1f70a |
2.24 GB | b05296f054d16133β¦ |
pi05_abl10k_dyn_v1_best/params/ocdbt.process_0/d/58cc5f3431155fb30201b1702d7b97ad |
2.24 GB | 002064c3ff1b935a⦠|
pi05_abl10k_release_v1_best/params/ocdbt.process_0/d/92f24ba7c6a954ce626c3de1148acbb6 |
2.24 GB | 7f19d4bff8d38769β¦ |
pi05_abl10k_ref_v1_best/params/ocdbt.process_0/d/21834220bef19e7a1b90af86ef0af4b5 |
2.24 GB | c28bc6eba076e0d8β¦ |
pi05_abl10k_ref_v1_best/params/ocdbt.process_0/d/936949880ab37500cdbcf9a92a411c35 |
2.24 GB | 91448d5d93641498β¦ |
pi05_abl10k_dyn_v1_best/params/ocdbt.process_0/d/c9cc30f2eb10389c33ee130c96fba845 |
2.24 GB | df57500d0b14c070β¦ |
| β¦ 82 smaller files | see SHA256SUMS |