Michi-Tsubaki/nextage_forceps_act

NEXTAGE Forceps ACT

ACT policy checkpoints for the NEXTAGE task of handing forceps to a hand.

Training

• Dataset: Michi-Tsubaki/hand_over_the_forceps_to_the_hand
• Architecture: ACT (Action Chunking with Transformers)
• Training split: first 45 of 50 episodes
• Training steps: 5,000
• Checkpoint interval: 1,000 steps
• Action chunk size: 60
• Batch size: 256
• Image inputs: top and right-wrist cameras, resized to 360 x 480
• State/action dimensions: 17
• Vision backbone: ResNet-18 with ImageNet pretrained weights

The root model.safetensors and config.json are the final checkpoint at step 5,000. The complete checkpoint history is available in checkpoint_001000 through checkpoint_005000. stats.npz contains the dataset statistics used by the training and evaluation pipeline, and train_args.json records the complete training arguments.

Loading the final checkpoint

from huggingface_hub import snapshot_download
from lerobot.policies.act.modeling_act import ACTPolicy

checkpoint_dir = snapshot_download("Michi-Tsubaki/nextage_forceps_act")
policy = ACTPolicy.from_pretrained(checkpoint_dir)

To load an intermediate checkpoint, pass its downloaded subdirectory to ACTPolicy.from_pretrained, for example checkpoint_003000.