manavgoel4/pi05_boxcutter_tray_camfixed_b16_h50_a25

Pi0.5 Boxcutter Tray Fine-Tune

Fine-tuned Pi0.5 model for the boxcutter-to-tray manipulation task.

Base Model

• lerobot/pi05_base

Dataset

• manavgoel4/molmo_act2_boxcutter_tray

Training Setup

• Policy: pi05
• Batch size: 16
• Steps: 30,000
• Checkpoint interval: 5,000
• chunk_size: 50
• n_action_steps: 25
• Data frequency: 10 Hz
• Action execution horizon: 2.5 seconds
• Dtype: bfloat16
• Train expert only: true
• Freeze vision encoder: true
• Gradient checkpointing: true
• Compile model: false
• Normalization:
  • ACTION: MEAN_STD
  • STATE: MEAN_STD
  • VISUAL: IDENTITY

Available Checkpoints

• 005000: checkpoints/005000/pretrained_model/
• 010000: checkpoints/010000/pretrained_model/
• 015000: checkpoints/015000/pretrained_model/
• 020000: checkpoints/020000/pretrained_model/
• 025000: checkpoints/025000/pretrained_model/
• 030000: checkpoints/030000/pretrained_model/
• last: checkpoints/last/pretrained_model/

Default Checkpoint

The repo root is set to checkpoint last. Loading the repo directly loads that checkpoint.

Python:

from lerobot.policies.pi05 import PI05Policy

policy = PI05Policy.from_pretrained("manavgoel4/pi05_boxcutter_tray_camfixed_b16_h50_a25")

Loading a Specific Checkpoint

Python:

from huggingface_hub import snapshot_download
from lerobot.policies.pi05 import PI05Policy

repo_dir = snapshot_download("manavgoel4/pi05_boxcutter_tray_camfixed_b16_h50_a25")

policy = PI05Policy.from_pretrained(
    f"{repo_dir}/checkpoints/last/pretrained_model"
)