Audio Pouring

We propose to make use of audio vibration sensing with a deep neural network named PouringNet to predict the liquid height from the audio fragment during the robotic pouring task. PouringNet is trained on our collected real-world pouring dataset with multimodal sensing data, which contains more than 3000 recordings of audio, force feedback, video and trajectory data of the human hand that performs the pouring task. Each record represents a complete pouring procedure. We conduct several evaluations on PouringNet with our dataset and robotic hardware. The results demonstrate that our PouringNet generalizes well across different liquid containers, positions of the audio receiver, initial liquid heights and types of liquid, and facilitates a more robust and accurate audio-based perception for robotic pouring.

• Project website: https://lianghongzhuo.github.io/AudioPouring/
• Preprint: https://arxiv.org/abs/1903.00650
• Video: https://www.youtube.com/watch?v=Za8dDjGFE1k
• Contact: [email protected], [email protected]

Overview

Experiments

• Pouring with different cups

  

  The cups are marked as #1-6 from left to right. And only cup #1/2/3 present in the dataset, while others are not included.

• Pouring with different initial heights on cup #3.

• Pouring with different microphone positions.

Installation

1. Install Anaconda and PyTorch:

   conda upgrade --all
   conda create -n pouring python=2.7 numpy ipython matplotlib mayavi yaml lxml seaborn
   conda activate pouring
   conda install -c conda-forge librosa trimesh pyglet
   pip install rospkg tensorboardx pyassimp==4.1.3
   # cpu:
   conda install tensorflow
   conda install pytorch-cpu torchvision-cpu -c pytorch
   # gpu:
   conda install tensorflow-gpu
   conda install pytorch torchvision cudatoolkit=9.0 -c pytorch

2. Clone and install this repository:

   git clone https://github.com/lianghongzhuo/AudioPouring.git
   cd AudioPouring
   AUDIO_POURING_DIR=${PWD}
   cd audio_pouring
   python setup.py develop

3. Install dependencies:

   • Install portaudio dependencies: according to this, install packages in this order will not remove any other packages

     sudo apt install libsndfile1-dev
     sudo apt install libjack-jackd2-dev
     sudo apt install portaudio19-dev

   • Make sure your current user name is in audio group
   • Other dependencies (only for robot experiment):

     cd ${AUDIO_POURING_DIR}
     sh audio_pouring_install.sh

4. Install following required ROS packages:

• portaudio_transport

Run demo

1. Bring up audio publishing node:

   roslaunch portaudio_transport publish.launch

2. Bring up a scale to get the ground truth height, if you do not have a ROS based scale, directly go to step 4.

3. Run demo code

   cd ${AUDIO_POURING_DIR}/audio_pouring
   python demo.py --cuda --bottle=1 --cavity-height=50

4. (In case when a ROS-based scale is not available) you can also use a normal scale and check the pouring result with the code below:

   from audio_pouring.utils.utils import weight2height
   print(weight2height(cup_id="1", cur_weight=0.02))

Network Training

1. Data preparation: generate a ~4s segment from a whole pouring sequence (pickle files):

   cd ${AUDIO_POURING_DIR}/audio_pouring/model
   python long_preprocess.py train mt
   python long_preprocess.py test mt

2. Data preparation: generate npy file list from that segment

   cd ${AUDIO_POURING_DIR}/audio_pouring/utils
   python generate_npy_list.py

3. Network training

   cd ${AUDIO_POURING_DIR}/audio_pouring
   python main_lstm.py --fixed --cuda --gpu=0 --bottle-train=0 --lstm --bs=32
   #args:
   #--fixed       : the input audio length is fixed (must set)
   #--lstm        : set to use lstm or gru
   #--bs          : set batch size
   #--bottle-train: set bottle id, if set to 0, then all the data are used

Generate your own bottle config file

• create a bottle config csv file and put it at ${AUDIO_POURING_DIR}/audio_pouring/config/bottles
• modify and run the code below:

  cd ${AUDIO_POURING_DIR}/audio_pouring/utils
  python generate_bottle_config.py

Dataset

• Containing video, audio, force/torque and position information collected during human pouring.
• We plan to release our dataset progressively. Now the audio part is available at Google Drive.

Citation

If you find this paper and code useful in your research, please consider citing:

@article{liang2019AudioPouring,
  title={Making Sense of Audio Vibration for Liquid Height Estimation in Robotic Pouring},
  author={Liang, Hongzhuo and Li, Shuang and Ma, Xiaojian  and Hendrich Norman and Gerkmann Timo and Zhang, Jianwei},
  journal={arXiv preprint arXiv:1903.00650},
  year={2019}
}