Chengbo Yuan, Chuan Wen, Tong Zhang, Yang Gao.
[Project Website] [Arxiv] [BibTex]
General Flow is a scalable framework designed to transfer low-level human skills to robots through actionable flow representation. By training exclusively on RGBD human video datasets, this approach successfully achieves an 81% success rate in zero-shot skill transfer across 18 tasks in 6 different scenes.
conda create -n gflow python=3.9
conda activate gflow
pip install -r requirement.txt
conda install -y pytorch=1.10.1 torchvision cudatoolkit=11.3 -c pytorch -c nvidia
pip install torch-scatter -f https://data.pyg.org/whl/torch-1.10.1+cu113.html
Then, follow the install.sh script to install the PointNeXt dependencies.
First conduct:
mkdir log
mkdir log/kpst_hoi4d
then download the pretrained model into log/kpst_hoi4d:
| Model | Eval-ADE | Eval-FDE | Param(M) | url |
|---|---|---|---|---|
| ScaleFlow-S | 0.0374 | 0.0501 | 0.906 | link |
| ScaleFlow-B | 0.0358 | 0.0477 | 5.622 | link |
| ScaleFlow-L | 0.0355 | 0.0470 | 17.088 | link |
Currently, we only guarantee the execution performance of specific tasks and instructions that are present in our training datasets. Please refer to metadata_stat.json for details. Instructions should be formatted as "something (action) Something (object)". For example, for the "open" action and "safe" object, the instruction should be formatted as "open Safe".
(1) download demo.tar.gz.
(2) create tool_repos/ folder, then install FastSAM in tool_repos/.
The final structure of the directory tree is:
- demo/
- input/
- safe_0_hand/
- output/
- tool_repos/
- FastSAM/
- aff_exec.py
Then, execute the Python script for interactive flow prediction:
CUDA_VISIBLE_DEVICES=0 python aff_exec.py --input_dir demo/input/safe_0_hand --desc open_Safe --pretrained_path log/kpst_hoi4d/ScaleGFlow-B/checkpoint/ckpt_best_train_scalegflow_b.pth
where --desc is the instruction (replace spaces with underscores for convenience). This command opens an interactive window that provides an interface to specify the segmentation prompt points and the 2D position of the gripper, as shown in the figure below.
The execution results, along with the visualization file, will be saved in demo/output/robot_exec.pkl. Subsequently, run python vis_exec.py -r demo/output/robot_exec.pkl for visualization.
Additionally, it's possible to directly specify the gripper position and segmentation prompt. For more information, see aff_exec.py.
The collected RGBD videos of "fold Clothes" task execution could be downloaded from link.
The HOI4D pipeline is located in the data/HOI4D directory. First, navigate to this directory by entering:
cd data/HOI4D
Next, execute the following command to extract the general flow labels:
CUDA_VISIBLE_DEVICES=0 python label_gen.py --data_root $HOI4D_ROOT --anno_root $HOI4D_ANNO_ROOT --idx_file $HOI4D_RELASE_FILE --output_root $SAVE_ROOT
In this command, $HOI4D_ROOT refers to the root directory that stores the RGBD data for the HOI4D dataset, HOI4D_ANNO_ROOT is the root directory where the annotations are stored, HOI4D_RELASE_FILE is the file path to the release.txt of the HOI4D dataset and $SAVE_ROOT is the root directory where the extracted general flow dataset will be saved.
After you have successfully extracted the general flow labels, proceed to generate the final dataset and the training split by running:
CUDA_VISIBLE_DEVICES=0 python label_gen_merge.py --output_root $SAVE_ROOT
You may check out HOI4D Official Repository, label_gen.py and label_gen_merge.py for more details.
The Raw RGBD Video pipeline is located in the data/RVideo directory. First, navigate to this directory by entering:
cd data/RVideo
Then install CLIP, FastSAM, 100DoH and CoTracker in tool_repos/ directory.
Next, execute the following command to extract the general flow labels:
CUDA_VISIBLE_DEVICES=0 python label_gen_demo.py --raw_data_root $Video_ROOT --save_root $SAVE_ROOT
In this command, $Video_ROOT refers to the root directory that stores the RGBD video data and $SAVE_ROOT is the root directory where the extracted general flow dataset will be saved.
After you have successfully extracted the general flow labels, you may merge the result into the HOI4D extraction dataset by running:
CUDA_VISIBLE_DEVICES=0 python merge_egosoft2hoi4d.py --hoi4d_dir $HOI4D_FLOW_DIR --egosoft_dir $RVideo_FLOW_DIR
where $HOI4D_FLOW_DIR and $RVideo_FLOW_DIR refers to the root directory that stores the extraction result of HOI4D dataset and Raw RGBD Videos. You may check out label_gen_demo.py and merge_egosoft2hoi4d.py for more details.
To start modeling training on the extracted dataset:
CUDA_VISIBLE_DEVICES=0 python train.py --cfg cfg/kpst_hoi4d/EarlyRegPNX-b.yaml val_save=80 --commit AD_base
The pre-process of datasets (for training acceleration) will be performed first (only be performed once). Then the training process will be started automatically.
To get prediction and evaluation on the extracted dataset:
CUDA_VISIBLE_DEVICES=0 python inference.py -i $IDX -n $SAMPLE -p $CKPT
where $IDX represents the index of the data point in the test set, $SAMPLE is the number of trajectories to sample, and $CKPT is the filepath of the model weight (.pth file). This command will evaluate the test set of the General Flow Datasets (HOI4D + Clothes) and save the visualization file in results/HOI4D/$model_name/$IDX_n$SAMPLE.pkl. For visualization, you can run python vis_inference.py -r results/HOI4D/$model_name/$IDX_n$SAMPLE.pkl.
This repository is based on the code from PointNeXt, HOI4D, FastSAM, 100DoH, CoTracker, PerAct and Deoxys.
If you find this repository useful, please kindly acknowledge our work :
@article{yuan2024general,
title={General Flow as Foundation Affordance for Scalable Robot Learning},
author={Yuan, Chengbo and Wen, Chuan and Zhang, Tong and Gao, Yang},
journal={arXiv preprint arXiv:2401.11439},
year={2024}
}