Ο0.5 β Relocating a block
This is an independent SFT policy initialized from the official OpenPI
pi05_base pretrained checkpoint. It uses 2 recorded real-robot execution
episodes for this task from Liukaikai/real-tasks at revision
de6347b6b9bf603a2578725b48def897f2d3b2bf. The four policies use 8 episodes in total.
Access requests require manual approval by the model owner.
Checkpoint format
params/ contains native OpenPI Orbax parameters. assets/ contains the
task-specific normalization statistics. model_config.json contains the model
and action-interface configuration. training_recipe.json records the
action representation, observations, optimizer settings, and training budget.
checkpoint_validation.json records checkpoint restoration and finite-action
inference verification. Any packaged Python inference helpers accompany this
checkpoint and should be used with the OpenPI version recorded in the recipe.
Training and validation losses describe the supplied demonstration data. Robot task-success evaluation is pending.
Offline loading and inference
Request access on this model page and wait for the owner's approval. In an environment with OpenPI installed, authenticate and download the checkpoint.
hf auth login
hf download Liukaikai/g2e-pi05-real-task2 --local-dir ./checkpoint
The training source records OpenPI commit fa431ecd893ee706883e64fe5fe1464ec8cd928d and individual source
file hashes in training_recipe.json. Use that OpenPI environment for checkpoint
compatibility. The packaged inference_policy.py and model_adapter.py provide
the model-specific observation and action transforms.
Inputs contain the 16 measured joint/motor positions in this exact order.
| Indices | Values |
|---|---|
| 0β6 | Left-arm joint positions |
| 7β13 | Right-arm joint positions |
| 14 | Left-gripper motor position |
| 15 | Right-gripper motor position |
The recorded arm and gripper positions use raw q values in radians. Actions
are absolute targets in the same order and units. Observation images are
RGB uint8 arrays of shape (224, 224, 3), resized with preserved aspect ratio
and centered black padding. The three image keys are images_base for the
Orbbec camera, images_left for the left wrist, and images_right for the right
wrist. The nominal recording rate is 10 Hz and the prediction horizon is
32 samples.
The example below loads one observation from a locally prepared episode NPZ with these fields. Model normalization and the internal padding to 32 action dimensions are handled by the packaged helpers.
import json
import sys
from pathlib import Path
import numpy as np
checkpoint = Path("./checkpoint").resolve()
sys.path.insert(0, str(checkpoint))
from inference_policy import RecordedTaskPolicy
recipe = json.loads((checkpoint / "training_recipe.json").read_text())
policy = RecordedTaskPolicy(checkpoint, recipe["task"], recipe["prompt"])
with np.load("/path/to/prepared_episode.npz", allow_pickle=False) as episode:
frame = 0
state = episode["state"][frame].astype(np.float32)
images = {key: episode[key][frame] for key in
("images_base", "images_left", "images_right")}
assert state.shape == (16,)
actions = policy.infer(state, images, num_steps=10)
assert actions.shape == (32, 16)
assert np.isfinite(actions).all()
print(actions.shape)
This example performs offline inference. The deployment application supplies synchronized observations and converts the returned targets to its calibrated robot-control interface.