You can deploy this model and start earning money today!

pravsels/
pi05-so101-stacking-magnetic-cubes-isambard

Name: pi05-so101-stacking-magnetic-cubes-isambard
Availability: InStock

Views

No views yet

pi0.5 SO101 Stacking Magnetic Cubes (Isambard)

Fine-tuned pi0.5 checkpoint for SO101 magnetic-cube stacking. Trained on Isambard GH200 (4× GPU node). Training is in progress — this repo publishes interim checkpoints under step_<N>/params/; final 50k may be added as step_49999/params/.

Published checkpoints

Directory	Step	Loss (approx)	Notes
`step_10000/params/`	10,000	0.0179	Interim; root `MODEL_PASSPORT.json` / `SIGNOFF.json`
`step_15000/params/`	15,000	0.0151	Interim; passport/signoff under `step_15000/`

Each step directory may include its own MODEL_PASSPORT.json and SIGNOFF.json when published separately from the root (10k) package.

Experiment

Objective: Fine-tune pi0.5 on lorenzouttini/so101_stacking_magnetic_cubes.
Weight init: pi0.5 base (pi05_base).
Training target: 50,000 steps (save_interval=5000).

Config

Config name: pi05_so101_stacking_magnetic_cubes
Experiment name: so101_stacking_magnetic_cubes
Model: pi0.5 (pi05=True, action_horizon=30)
Batch size: 32 · LR: 2.5e-5 cosine (1k warmup) · EMA: 0.999
Delta actions: yes · Default prompt: stack the magnetic cubes

Dataset

lorenzouttini/so101_stacking_magnetic_cubes

W&B

Training run (synced partial; full 50k curves after final sync)

Repo structure

README.md
TRAINING_LOG.md
MODEL_PASSPORT.json          # signed for step_10000 package layout
SIGNOFF.json
assets/                      # norm stats, valid_indices, reference_test_vector
step_10000/params/           # interim checkpoint (params only)
step_15000/params/           # interim checkpoint (params only)
step_15000/MODEL_PASSPORT.json
step_15000/SIGNOFF.json

Usage

Point checkpoint_path at the published step directory:

python

from openpi.training.config import get_config
from openpi.serving.policy_server import PolicyServer

config = get_config("pi05_so101_stacking_magnetic_cubes")
server = PolicyServer(config, checkpoint_path="step_10000/params")

Verify integrity before load:

validate-checkpoint . --require-signoff

pravsels/pi05-so101-stacking-magnetic-cubes-isambard