Learning Manipulation by Predicting Interaction (MPI)

Project Website | Paper | RSS 2024

🔥 Highlight

​MPI is an interaction-oriented representation learning method towards robot manipulation:

• Instruct the model towards predicting transition frames and detecting manipulated objects with keyframes.
• Foster better comprehension of “how-to-interact” and “where-to-interact”.
• Acquire more informative representations during pre-training and achieve evident improvement across downstream tasks.

🎥 Demo

Real-world robot experiments on complex kitchen environment.

Take the spatula off the shelf (2x speed)	Lift up the pot lid (2x speed)
take_the_spatula_2.mp4	lift_up_the_pot_lid_2.mp4
Close the drawer (2x speed)	Put pot into sink (2x speed)
close_drawer_2.mp4	mov_pot_into_sink_2.mp4

🚀 News

• [2024/05/31] We released the implementation of pre-training and evaluation on Referring Expression Grounding task.
• [2024/06/04] We released our paper on arXiv.
• [2024/06/16] We released the model weights.
• [2024/07/05] We released the evaluation code on Franka Kitchen environment.

Getting Started

• Installation
• Get representation
• Pre-training
• Evaluation

Installation

Step 1. Clone and setup MPI dependency:

git clone https://github.com/OpenDriveLab/MPI
cd MPI
pip install -e .

Step 2. Prepare the language model, you may download DistillBERT from HuggingFace

Get representation

🧳 Checkpoints

To directly utilize MPI for extracting representations, please download our pre-trained weights:

Model	Checkpoint	Params.	Config
MPI-Small	GoogleDrive	22M	GoogleDrive
MPI-Base	GoogleDrive	86M	GoogleDrive

Your directory tree should look like this:

checkpoints
├── mpi-small
|   |—— MPI-small-state_dict.pt  
|   └── MPI-small.json
└── mpi-base
    |—— MPI-base-state_dict.pt    
    └── MPI-base.json

Obtain representation from pretrained MPI

We provide a example code get_representation.py to show how to obtain the pre-trained MPI features. The MPI encoder by default requires two images as input. In downstream tasks, we simply replicate the current observation to ensure compatibility.

The following diagram presents the composition and arrangement of the extracted tokens:

Pre-training

Prepare Pre-training Dataset

Download Ego4D Hand-and-Object dataset:

# Download the CLI
pip install ego4d
# Select Subset Of Hand-and-Object
python -m ego4d.cli.cli --output_directory=<path-to-save-dir> --datasets clips annotations  --metadata --version v2 --benchmarks FHO

Your directory tree should look like this:

$<path-to-save-dir>
├── ego4d.json
└── v2
    |—— annotations  
    └── clips

Preprocess dataset for pre-training MPI:

python prepare_dataset.py --root_path <path-to-save-dir>/v2/

Pre-training script

Pre-train MPI on 8 Nvidia A100 GPUs:

torchrun --standalone --nnodes 1 --nproc-per-node 8 pretrain.py

Evaluation

Referring Expression Grounding

Step 1. Prepare the OCID-Ref dataset following this repo. Then put the dataset to

./mpi_evaluation/referring_grounding/data/langref

Step 2. Initiate evaluation with

python mpi_evaluation/referring_grounding/evaluate_refer.py test_only=False iou_threshold=0.5 lr=1e-3 \
model=\"mpi-small\" \
save_path=\"MPI-Small-IOU0.5\" \
eval_checkpoint_path=\"path_to_your/MPI-small-state_dict.pt\" \
language_model_path=\"path_to_your/distilbert-base-uncased\" \

or you can simply use

bash mpi_evaluation/referring_grounding/eval_refer.sh

Franka Kitchen

Following the guidebook to setup Franka Kitchen environment and download the expert demonstrations.

Evaluating visuomotor control on Franka Kitchen environment with 25 expert demonstration.

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=mpi_evaluation/franka_kitchen/MPIEval/core python mpi_evaluation/franka_kitchen/MPIEval/core/hydra_launcher.py hydra/launcher=local hydra/output=local env="kitchen_knob1_on-v3" camera="left_cap2" pixel_based=true embedding=ViT-Small num_demos=25 env_kwargs.load_path=mpi-small bc_kwargs.finetune=false job_name=mpi-small seed=125 proprio=9

Citation

If you find the project helpful for your research, please consider citing our paper:

@article{zeng2024learning,
  title={Learning Manipulation by Predicting Interaction},
  author={Zeng, Jia and Bu, Qingwen and Wang, Bangjun and Xia, Wenke and Chen, Li and Dong, Hao and Song, Haoming and Wang, Dong and Hu, Di and Luo, Ping and others},
  journal={arXiv preprint arXiv:2406.00439},
  year={2024}
}

Acknowledgment

The code of this work is built upon Voltron and R3M. Thanks for their open-source work!