ReNeLib

ReNeLiB: Real-time Neural Listening Behavior Generation for Socially Interactive Agents

Teaser Image

Quick Links

Demo Video

Abstract

Flexible and natural nonverbal reactions to human behavior remain a challenge for socially interactive agents (SIAs), especially those animated using predominantly hand-crafted rules. Our toolkit introduces a novel approach to enhancing the real-time, machine learning-based, conversational behavior generation capabilities of SIAs, considering the challenging real-time requirements of human-agent interaction scenarios.

Real-time Feature Extraction: Captures multi-modal social cues from the user.
Behavior Generation: Utilizes a state-of-the-art neural network approach.
Visualization: Supports both FLAME-based and Apple ARKit-based interactive agents.

We also provide new pre-trained behavioral generation models based on real-life psychotherapy interactions, enabling domain-specific listening behaviors generation. Our toolkit will be publicly available, serving as a valuable resource for researchers.

Installation

Using Conda

conda env create python=3.8 --file environment.yml

Or alternatively:

conda env create python=3.8 --file conda_spec_file.txt

If any dependency installation fails, particularly pytorch3d and torch, install them after initiating the conda environment. Ensure to check your GPU’s CUDA compatibility and install the correct Nvidia drivers and CUDA toolkits. Follow the installation instructions for pytorch and pytorch3d accordingly.

Docker Container

Execute the following commands:

docker run --rm -it pytorch/pytorch:1.11.0-cuda11.3-cudnn8-runtime
$conda install -c fvcore -c iopath -c conda-forge fvcore iopath
$conda install pytorch3d=0.7.0 -c pytorch3d
$ipython
$from pytorch3d.structures import Meshes

Demo

Running the System

Module 1: Behaviour Generation

Navigate to behaviour_predictor/behaviour/online_with_webcam_microphone and run:
```
python behaviour_predictor_with_VAD.py
```
Module 2: FastAPI Backend

Navigate to IVA/fastApi_backend, ensure all FastAPI dependencies are installed, and start the server:
```
python fast_api_server.py
```
After starting the backend, connect the front end with Gloria to test the communication. Navigate to IVA/front-end and start index.html. Ensure your browser and JavaScript dependencies are up-to-date. Click the connect button to establish a connection to the backend once Gloria is visible on your web browser.
Module 3: FLAME Webcam Extractor

Navigate to extractors/src/feature_extraction_online and start:
```
python webcam_flame_extractor_simple_version.py
```
Then, start mfcc extractor:
```
python microphone_mfcc_extractor_with_VAD.py
```
These modules stream data to the behavior generator for behavior prediction.

Data

To access the therapy dataset and pre-trained models, please contact me via email: daksitha.withanage.don@uni-a-de or daksitha.withanage@gmail.com

Cite Our Paper

If you utilize our work, please cite our paper:

@inproceedings{10.1145/3577190.3614133,
  author = {Withanage Don, Daksitha Senel and M"{u}ller, Philipp and Nunnari, Fabrizio and Andr'{e}, Elisabeth and Gebhard, Patrick},
  title = {ReNeLiB: Real-Time Neural Listening Behavior Generation for Socially Interactive Agents},
  year = {2023},
  isbn = {9798400700552},
  publisher = {Association for Computing Machinery},
  address = {New York, NY, USA},
  url = {https://doi.org/10.1145/3577190.3614133},
  doi = {10.1145/3577190.3614133},
  pages = {507–516},
  numpages = {10},
  location = {Paris, France},
  series = {ICMI '23}
}

Acknowledgements

Special thanks to Radek Daněček and Evone Ng for their invaluable tips and code contributions.

For further collaborations or queries, feel free to contact me.