ivanovtech/pi05-a35LKYALAosS

OpenRoboto Gate 5t epoch 2

Complete merged PyTorch π0.5 checkpoint prepared for OpenRoboto subnet 80, round 1.

Provenance

• Official parent: 0xjulius/pi05-LZC4rWJGrdCn
• Parent revision: b2da1b725ed89d38a4e19b9df14f79fde9f2217e
• Training method: supervised corrective task-6 rehearsal with distribution-matched task-1 retention replay and a representative broad retention bank.
• Dense checkpoint SHA-256: b3c058226db50871b3e1cf3af2dc13bfbd4d13e8acf5b60d67e49f55cfe7c6cf
• Normalization stats SHA-256: 7f920edb4ab18f811c5f13bde40d2cb67bd95411ea89e7fc0bc9bdb02438e793

Local evaluation

Paired Gate 5t development evidence:

gate	parent	candidate
task-6 terminal success	0/15	10/15
task-6 correct object choice	0/15	14/15
task-1 terminal success	14/15	12/15
broad retention	24/24	23/24

An additional eight-pair random diagnostic tied 6/8 to 6/8, with zero paired wins and zero paired regressions. These local panels do not estimate the official full-benchmark score.

Format

• model.safetensors: complete merged PyTorch π0.5 weights, not a bare adapter.
• assets/physical-intelligence/libero/norm_stats.json: LIBERO normalization statistics required by the evaluator.
• Official openroboto check and the public evaluator's check_model.py both pass with no warnings or errors.

Limitations

The candidate is submitted to obtain a broad official measurement. It improved the targeted task-6 development panel but did not meet the frozen task-1 retention threshold by one success, and it has not yet run the official full benchmark.