Multimodal Diffusion Transformer: Learning Versatile Behavior from Multimodal Goals

Paper, Project Page, RSS 2024

Moritz Reuss¹, Ömer Erdinç Yağmurlu, Fabian Wenzel, Rudolf Lioutikov¹

¹Intuitive Robots Lab, Karlsruhe Institute of Technology

This is the official code repository for the paper Multimodal Diffusion Transformer: Learning Versatile Behavior from Multimodal Goals.

Pre-trained models are available here.

Performance

Results on the CALVIN benchmark (1000 chains):

Train	Method	1	2	3	4	5	Avg. Len.
D	HULC	82.5%	66.8%	52.0%	39.3%	27.5%	2.68±(0.11)
	LAD	88.7%	69.9%	54.5%	42.7%	32.2%	2.88±(0.19)
	Distill-D	86.7%	71.5%	57.0%	45.9%	35.6%	2.97±(0.04)
	MT-ACT	88.4%	72.2%	57.2%	44.9%	35.3%	2.98±(0.05)
	MDT (ours)	93.3%	82.4%	71.5%	60.9%	51.1%	3.59±(0.07)
	MDT-V (ours)	93.9%	83.8%	73.5%	63.9%	54.9%	3.70±(0.03)*
ABCD	HULC	88.9%	73.3%	58.7%	47.5%	38.3%	3.06±(0.07)
	Distill-D	86.3%	72.7%	60.1%	51.2%	41.7%	3.16±(0.06)
	MT-ACT	87.1%	69.8%	53.4%	40.0%	29.3%	2.80±(0.03)
	RoboFlamingo	96.4%	89.6%	82.4%	74.0%	66.0%	4.09±(0.00)
	MDT (ours)	97.8%	93.8%	88.8%	83.1%	77.0%	4.41±(0.03)
	MDT-V (ours)	99.1%	96.8%	92.8%	88.5%	83.1%	4.60±(0.05)*

*: 3.72±(0.05) (D) and 4.52±(0.02) (ABCD) in the paper. Performance is higher than reported given some fixes in the camera-ready code version.

Installation

To begin, clone the repository locally:

git clone --recurse-submodules [email protected]:intuitive-robots/mdt_policy.git
export MDT_ROOT=$(pwd)/mdt_policy

Install the requirements: (Note we provided a changed verison of pyhash, given numerous problems when installing it manually)

cd $MDT_ROOT
conda create -n mdt_env python=3.8
conda activate mdt_env
cd calvin_env/tacto
pip install -e .
cd ..
pip install -e .
cd ..
pip install setuptools==57.5.0
cd pyhash-0.9.3
python setup.py build
python setup.py install
cd ..

Next we can install the rest of the missing packages:

pip install -r requirements.txt

Evaluation

Step 1 - Download CALVIN Datasets

If you want to train on the CALVIN dataset, choose a split with:

cd $MDT_ROOT/dataset
sh download_data.sh D | ABCD

(Optional) Preprocessing with CALVIN

Since MDT uses action chunking, it needs to load multiple (~10) episode_{}.npz files for each inference. In combination with batching, this results in a large disk bandwidth needed for each iteration (usually ~2000MB/iteration). This has the potential of significantly reducing your GPU utilization rate during training depending on your hardware. Therefore, you can use the script extract_by_key.py to extract the data into a single file, avoiding opening too many episode files when using the CALVIN dataset.

Usage example:

python preprocess/extract_by_key.py -i /YOUR/PATH/TO/CALVIN/ \
    --in_task all

Params:

Run this command to see more detailed information:

python preprocess/extract_by_key.py -h

Important params:

• --in_root: /YOUR/PATH/TO/CALVIN/, e.g /data3/geyuan/datasets/CALVIN/
• --extract_key: A key of dict(episode_xxx.npz), default is 'rel_actions', the saved file name depends on this (i.e ep_{extract_key}.npy)

Optional params:

• --in_task: default is 'all', meaning all task folders (e.g task_ABCD_D/) of CALVIN
• --in_split: default is 'all', meaning both training/ and validation/
• --out_dir: optional, default is 'None', and will be converted to {in_root}/{in_task}/{in_split}/extracted/
• --force: whether to overwrite existing extracted data

Thanks to @ygtxr1997 for debugging the GPU utilization and providing a merge request.

Step 2 - Download Pre-trained Models

Name	Split	1	2	3	4	5	Avg. Len.	Model	Seed	eval: sigma-min
mdtv-1-abcd	ABCD -> D	99.6%	97.5%	94.0%	90.2%	84.7%	4.66	MDT-V	142	1.000
mdtv-2-abcd	ABCD -> D	99.0%	96.2%	92.4%	88.3%	82.5%	4.58	MDT-V	242	1.000
mdtv-3-abcd	ABCD -> D	98.9%	96.7%	92.1%	87.1%	82.1%	4.57	MDT-V	42	1.000
mdtv-1-d	D -> D	93.8%	83.4%	72.6%	63.2%	54.4%	3.67	MDT-V	142	0.001
mdtv-2-d	D -> D	94.0%	84.0%	73.3%	63.5%	54.2%	3.69	MDT-V	242	0.001
mdtv-3-d	D -> D	93.9%	84.0%	74.6%	65.0%	56.3%	3.74	MDT-V	42	1.000

You can find all of the aforementioned models under here.

Step 3 - Run

Adjust conf/mdt_evaluate.conf according to the model you downloaded. Important keys are:

• voltron_cache: set a rw path in order to avoid downloading Voltron from huggingface each run.
• num_videos: n first rollouts of the model will be recorded as a gif.
• dataset_path and train_folder: location of the downloaded files.
• sigma_min: use the values in the table for best results with a given model.

Then run:

python mdt/evaluation/mdt_evaluate.py

Training

To train the MDT model with the maximum amount of available GPUS, run:

python mdt/training.py

For replication of the originial training results I recommend to use 4 GPUs with a batch_size of 128 and train them for 30 (20) epochs for ABCD (D only). You can find the exact configuration used for each of our pretrained models in .hydra/config.yaml under the respective model's folder.

---

Acknowledgements

This work is only possible because of the code from the following open-source projects and datasets:

CALVIN

Original: https://github.com/mees/calvin License: MIT

BESO

Original: https://github.com/intuitive-robots/beso License: MIT

Voltron

Original: https://github.com/siddk/voltron-robotics License: MIT

OpenAI CLIP

Original: https://github.com/openai/CLIP License: MIT

HULC

Original: https://github.com/lukashermann/hulc License: MIT

Citation

If you found the code usefull, please cite our work:

@inproceedings{
    reuss2024multimodal,
    title={Multimodal Diffusion Transformer: Learning Versatile Behavior from Multimodal Goals},
    author={Moritz Reuss and {\"O}mer Erdin{\c{c}} Ya{\u{g}}murlu and Fabian Wenzel and Rudolf Lioutikov},
    booktitle={Robotics: Science and Systems},
    year={2024}
}