Hydra: A Navigation World Action Model

Hydra is a Navigation World Action Model (WAM) with Discrete Latent Planning (DLP) and continuous Flow Matching Execution. You can find the source code, training, and deployment scripts in our GitHub Repository.

Download Pretrained Models

Download and unzip the weights in a folder. Pass the address to this folder in the deployment config file.

Model Name	Parameters	Download Link
Hydra	143.7M	📥 download .zip
VertiFormer	27.11M	📥 download .zip

How to Load via Python

You can load the .pth model from the table above directly into your script using the following snippet:

import torch
import urllib.request
import os

# 1. Define the direct download URL and local filename
MODEL_URL = "https://huggingface.co/mhnazeri/Hydra/blob/main"
local_weights_path = "hydra.pth"

# 2. Download the file if it doesn't exist locally
if not os.path.exists(local_weights_path):
    print("Downloading model weights...")
    urllib.request.urlretrieve(MODEL_URL, local_weights_path)

# 3. Load the weights into your PyTorch model
model = Hydra(cfg)  # load the yaml file and pass it to the model.
state_dict = torch.load(local_weights_path, map_location=torch.device('cpu'))
# model.load_state_dict(state_dict)
print("Weights loaded successfully!")

Fast Loading via huggingface_hub

If you have the huggingface_hub library installed (pip install huggingface_hub), you can fetch the file instantly:

import torch
from huggingface_hub import hf_hub_download

# This automatically downloads and caches the file locally
weights_path = hf_hub_download(repo_id="mhnazeri/hydra", filename="hydra.pth")

# Load directly into PyTorch
state_dict = torch.load(weights_path, map_location="cpu")