pi0.5 robot_only: 30000-step EMA checkpoint
Private research delivery for maple0617/pi05-robot32-30k. This package is intended
for the user's authorized private repository; it does not change its visibility.
Training and contents
This experiment was independently initialized from the same pinned pi05_base
checkpoint. It used full fine-tuning for 30000 optimizer steps, a global batch of
32 (0 ego + 32 robot), and EMA decay 0.999.
The exact model, optimizer, data revisions, source hashes, normalization statistics
and training recipe are recorded in recipe.json and recipe_assets/.
30000/params/ contains the complete EMA inference parameter item in its original
Orbax/JAX layout, including metadata and data shards. 30000/assets/ contains
checkpoint assets. Both recipe_assets/robot/ and recipe_assets/ego/ are retained.
metrics.jsonl and status.json preserve the experiment's original records.
Raw training logs, training datasets, image caches, earlier checkpoints and the
optimizer/training-state checkpoint item are not included. This package alone
cannot resume the original training state.
Verification and provenance
Both training completion and the final full checkpoint verification passed. The original verification fully read all 51 EMA arrays and all 156 train-state arrays and checked that their values were finite. The full training-state item was checked on the source server but is not included in this weight delivery.
The unmodified proof is fsdp4_production_20260928T231638Z_e15e494a-robot_only-30000-full.json.
Its SHA256 is 147d001154d0098fba5b4dbef3128b6ea63d98be6c951797a4ecec5362a99a9d.
Recipe, status, metrics, statistics and proof bytes are preserved without
redaction for this authorized private research handoff. They contain source
filesystem paths and experiment infrastructure identifiers. Source paths in
these records describe the training environment, not paths the recipient must use.
Preparation checks payload file counts and byte totals and uses source hard links; it does not compute new SHA256 hashes for the large weight files. Any subsequent transport manifest should distinguish transport hashes from the original proof.
Inference limitations
These checks establish checkpoint readability and finite values; they do not establish inference correctness, task success rates or real-robot deployment. Use matching custom OpenPI code and configuration plus verified input and action adapters. This is not a validated plug-and-play Transformers model or default OpenPI policy-server deployment.
Training uses no measured state input, a 50-target action horizon, and 14 effective
action dimensions padded to 32. For each hand, translation and rotation are
relative to action[0]; gripper values remain absolute. Robot inference must use
the matching normalization statistics, camera mapping, action representation,
coordinate reconstruction and control interface. Those adapters and real-robot
execution have not been validated by this checkpoint-only release.
Upstream terms and modification notice
This checkpoint modifies the pinned upstream pi0.5 base through the experiment
described above; the distributed model parameter files are fine-tuned derivatives.
The upstream OpenPI Apache 2.0 license is included as LICENSE. Gemma-derived
components are also subject to LICENSE_GEMMA.txt and NOTICE; the entire model
is not represented as exclusively Apache-2.0 licensed.
Use and distribution of the Gemma-derived model are subject to the restrictions in Section 3.2 of the included Gemma Terms of Use, including the Gemma Prohibited Use Policy. Recipients must comply with these incorporated restrictions and applicable terms. No rights to redistribute the source training datasets are granted by this package.