nock0912/aloha_act_policy

ACT Policy for ALOHA Simulation

This policy was trained using Imitation Learning (behavior cloning) on the ALOHA simulation dataset.

Model Details

• Policy Type: ACT (Action Chunking Transformer)
• Architecture: CNN encoder + Transformer decoder
• State Dimension: 14 (bimanual robot joints)
• Action Dimension: 14
• Action Chunk Size: 50
• Hidden Dimension: 512
• Transformer Layers: 6
• Training Epochs: 6
• Best Training Loss: 0.0008

Training Dataset

• Dataset: nock0912/aloha_sim_cube_dataset
• Environment: gym_aloha/AlohaTransferCube-v0
• Task: Transfer cube from one location to another

Usage

import torch
from train_act_policy import SimpleACTPolicy

# Load model
model = SimpleACTPolicy(
    state_dim=14,
    action_dim=14,
    hidden_dim=512,
    chunk_size=50,
    n_heads=8,
    n_layers=6
)
model.load_state_dict(torch.load("model.pt"))
model.eval()

# Inference
with torch.no_grad():
    actions = model(images, states)  # Returns 50 future actions

Citation

If you use this policy, please cite the LeRobot project.