Model Card for act-openarm-2-cell-pick_up_cube_mujoco

Action Chunking with Transformers (ACT) policy trained with LeRobot on the k1000dai/openarm-2-cell-pick_up_cube_mujoco-lerobot dataset: a bimanual OpenArm robot performing a cube pick-up task in MuJoCo simulation.

See the full LeRobot documentation at huggingface.co/docs/lerobot.

Model Details

• Policy type: ACT (Action Chunking with Transformers)
• Vision backbone: ResNet-18 (ImageNet pretrained)
• Chunk size / action steps: 100 / 100
• VAE: enabled (latent dim 32, KL weight 10.0)
• Transformer: dim_model 512, 8 heads, 4 encoder layers, 1 decoder layer

Inputs

Feature	Type	Shape
observation.state	STATE	(16,)
observation.images.ceiling	VISUAL	(3, 600, 960)
observation.images.head_left	VISUAL	(3, 720, 1280)
observation.images.head_right	VISUAL	(3, 720, 1280)
observation.images.wrist_left	VISUAL	(3, 600, 960)
observation.images.wrist_right	VISUAL	(3, 600, 960)

Outputs

Feature	Type	Shape
action	ACTION	(16,)

Training

• Dataset: k1000dai/openarm-2-cell-pick_up_cube_mujoco-lerobot
• Steps: 10,000
• Batch size: 8
• Optimizer: AdamW (lr 1e-5, weight decay 1e-4, grad clip 10.0)
• Normalization: mean/std for visual, state, and action features
• Seed: 1000

How to Use

from lerobot.policies.act.modeling_act import ACTPolicy

policy = ACTPolicy.from_pretrained("k1000dai/act-openarm-2-cell-pick_up_cube_mujoco")

To evaluate or fine-tune with the LeRobot CLI, pass --policy.path=k1000dai/act-openarm-2-cell-pick_up_cube_mujoco.