Download Checkpoints#

We provide pre-trained model checkpoints for all four benchmark tasks, hosted on Hugging Face. These checkpoints allow you to immediately evaluate mindmap performance without training your own models.

Available Checkpoints#

mindmap Checkpoints: nvidia/PhysicalAI-Robotics-mindmap-Checkpoints

Prerequisites#

Before downloading, ensure you have:

Hugging Face account registered
Read-access token ready for authentication

Download Instructions#

Install the Hugging Face Hub CLI:
```
pip install -U "huggingface_hub[cli]"
```
Authenticate with your token:
```
hf auth login
```

Download the checkpoints:

hf download \
    nvidia/PhysicalAI-Robotics-mindmap-Checkpoints \
    --local-dir <LOCAL_CHECKPOINT_PATH>

Checkpoint Structure#

The downloaded repository contains a folder for each task with the following structure:

📂 <LOCAL_CHECKPOINT_PATH>
├── 📂 cube_stacking_checkpoint
│   ├── best.pth
│   ├── last.pth
│   └── training_args.json
├── 📂 mug_in_drawer_checkpoint
│   ├── ...
├── 📂 drill_in_box_checkpoint
│   ├── ...
├── 📂 stick_in_bin_checkpoint
│   ├── ...
└── README.md

For detailed information about the checkpoint structure, see Checkpoint Structure.

Note

Replace <LOCAL_CHECKPOINT_PATH> with your desired local directory path for storing the checkpoints.