Octo-Small PyTorch Model

This is the octo-small model converted to PyTorch format.

Model Description

Octo is a generalist robot policy trained on diverse robot manipulation tasks.

• Paper: Octo: An Open-Source Generalist Robot Policy
• Original Implementation: octo-models/octo
• Model Size: octo-small

Usage

Loading the pretrained model

import torch
from safetensors.torch import load_file
import json

# Load config
with open('config.json', 'r') as f:
    config = json.load(f)

# Initialize model (assuming you have the model class)
from octo_pytorch.model.octo_model import OctoModel
model = OctoModel(model_name=config['model_name'], repeat_task_tokens=config['repeat_task_tokens'])

# Load weights (T5 encoder weights will be loaded automatically from HuggingFace Hub)
state_dict = load_file('model.safetensors')
model.load_state_dict(state_dict, strict=False)  # strict=False because T5 weights are not in the file

Note: The T5-base language encoder weights are not included in this upload to save space. They will be automatically downloaded from HuggingFace Hub when you initialize the model.

Model Architecture

• Transformer: 12 layers, 384 dim, 6 heads
• Vision Encoder: Custom CNN (SmallStem16)
• Language Encoder: T5-Base
• Action Head: Diffusion policy with 4 action steps
• Max Horizon: 10 timesteps
• Action Dimension: 7

Files

• model.safetensors: Model weights in safetensors format
• config.json: Model configuration
• dataset_statistics.npy: Dataset statistics used for normalization (if available)

Citation

If you use this model, please cite:

@article{octo_2023,
    title={Octo: An Open-Source Generalist Robot Policy},
    author={Octo Model Team et al.},
    journal={arXiv preprint arXiv:2405.12213},
    year={2024}
}