pi05-table-bussing-base-policy-20k

pi05-table-bussing-base-policy-20k

Model checkpoint for step 20000.

Checkpoint Information

• Checkpoint Path: 20000
• Parent Directory: pi05_table_bussing_sanity
• Full Path: /home/prakhar-goel/data_farm/checkpoints/pi05/pi05_table_bussing_sanity/20000

Structure

This checkpoint contains:

• params/: Model parameters in JAX/Flax format
• train_state/: Training state including optimizer states
• assets/: Normalization statistics and other training assets
• _CHECKPOINT_METADATA: Checkpoint metadata

Loading the Checkpoint

To load this checkpoint, use:

from pathlib import Path
import jax
import orbax.checkpoint as ocp
from etils import epath

checkpoint_dir = Path("path/to/downloaded/checkpoint")

# Load parameters
restored = ocp.CheckpointManager(checkpoint_dir).restore(
    checkpoint_dir,
    items={
        "params": ...,
        "train_state": ...,
    }
)

Or use with OpenPI/LeRobot policy loading:

from openpi.policies.policy_config import create_trained_policy

# Load policy from checkpoint
policy = create_trained_policy(
    train_config=your_config,
    checkpoint_dir=str(checkpoint_dir)
)

Metadata

{
  "item_handlers": {
    "assets": "openpi.training.checkpoints.CallbackHandler",
    "params": "orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler",
    "train_state": "orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler"
  },
  "metrics": {},
  "performance_metrics": {},
  "init_timestamp_nsecs": 1761916229895447133,
  "commit_timestamp_nsecs": 1761916301719153113,
  "custom_metadata": {}
}