🦾 Heterogenous Pre-trained Transformers

Lirui Wang, Xinlei Chen, Jialiang Zhao, Kaiming He

Neural Information Processing Systems (Spotlight), 2024

---

This is a pytorch implementation for pre-training Heterogenous Pre-trained Transformers (HPTs). The pre-training procedure train on mixture of embodiment datasets with a supervised learning objective. The pre-training process can take some time, so we also provide pre-trained checkpoints below. You can find more details on our project page. An alternative clean implementation of HPT in Hugging Face can also be found here.

TL;DR: HPT aligns different embodiment to a shared latent space and investigates the scaling behaviors in policy learning. Put a scalable transformer in the middle of your policy and don’t train from scratch!

⚙️ Setup

1. pip install -e .

Install (old-version) Mujoco

mkdir ~/.mujoco
cd ~/.mujoco
wget https://mujoco.org/download/mujoco210-linux-x86_64.tar.gz  -O mujoco210.tar.gz --no-check-certificate
tar -xvzf mujoco210.tar.gz

# add the following line to ~/.bashrc if needed
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:${HOME}/.mujoco/mujoco210/bin
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/lib/nvidia
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64
export MUJOCO_GL=egl

🚶 Usage

0. Check out quickstart.ipynb for how to use the pretrained HPTs.
1. python -m hpt.run train policies on each environment. Add +mode=debug for debugging.
2. bash experiments/scripts/metaworld/train_test_metaworld_1task.sh test test 1 +mode=debug for example script.
3. Change train.pretrained_dir for loading pre-trained trunk transformer. The model can be loaded either from local checkpoint folder or huggingface repository.
4. Run the following scripts for mujoco experiments.

Metaworld 20 Task Experiments

bash experiments/scripts/metaworld/train_test_metaworld_20task_finetune.sh hf://liruiw/hpt-base

5. See here for defining and modifying the hyperparameters.
6. We use wandb to log the training process.

🤖 Try this On Your Own Dataset

0. For training, it requires a dataset conversion convert_dataset function for packing your own datasets. Check this for example.
1. For evaluation, it requires a rollout_runner.py file for each benchmark and a learner_trajectory_generator evaluation function that provides rollouts.
2. If needed, modify the config for changing the perception stem networks and action head networks in the models. Take a look at realrobot_image.yaml for example script in the real world.
3. Add dataset.use_disk=True for saving and loading the dataset in disk.

💽 Checkpoints

You can find pretrained HPT checkpoints here. At the moment we provide the following model versions:

Model	Size
HPT-XLarge	226.8M Params
HPT-Large	50.5M Params
HPT-Base	12.6M Params
HPT-Small	3.1M Params
HPT-Large (With Language)	50.6M Params

---

💾 File Structure

├── ...
├── HPT
|   ├── data            # cached datasets
|   ├── output          # trained models and figures
|   ├── env             # environment wrappers
|   ├── hpt             # model training and dataset source code
|   |   ├── models      # network models
|   |   ├── datasets    # dataset related
|   |   ├── run         # transfer learning main loop
|   |   ├── run_eval    # evaluation main loop
|   |   └── ...
|   ├── experiments     # training configs
|   |   ├── configs     # modular configs
└── ...

🕹️ Citation

If you find HPT useful in your research, please consider citing:

@inproceedings{wang2024hpt,
author    = {Lirui Wang, Xinlei Chen, Jialiang Zhao, Kaiming He},
title     = {Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers},
booktitle = {Neurips},
year      = {2024}
}

Contact

If you have any questions, feel free to contact me through email (liruiw@mit.edu). Enjoy!