CoBaIR is a python library for Context Based Intention Recognition. It provides the means to infer an intention from given context. An intention is a binary value e.g. repair pipe that can either be present or not. Only one intention can be present at a time. Context on the otherhand can have multiple discrete instantiations e.g. weather:sunny|cloudy|raining. If context values are continuous, discretizer functions can be used to create discrete values. From the inferred intention in a HRI scenario the robot can perform corresponding actions to help the human with a specific task.

Publications

For a more in-depth explanation consult the following papers:

• Concept Paper
• Full Paper for IEEE RO-MAN Conference

Install

CoBaIR needs Python >= 3.10, the web GUI needs Python >= 3.11. Install it with the extras you need:

pip install "cobair[all]"

You can install the library from your local copy after cloning this repo with pip using pip install ".[all]" or install the newest unreleased features from the main branch with pip install "cobair[all] @ git+https://github.com/dfki-ric/CoBaIR.git@main"

Usage

from cobair.model import Config
from cobair.infer import Inference

config = Config.load('small_example.yml')
inference = Inference(config)
intention, decision_threshold, probabilities = inference.infer({'speech commands': 'pickup'})

cobair.model.Config holds the scenario and has methods to edit, validate, load and save it. cobair.infer.Inference infers the intentions for given evidence. See the API documentation.

The 4.x API (from CoBaIR.bayes_net import BayesNet, load_config) still works until 6.0.0, but raises a DeprecationWarning.

Known Issues

On some Linux Distros there seems to be a problem with a shared library. This Solutions suggests to export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libstdc++.so.6 which works on Ubuntu 22.04.

Web GUI

The web GUI is the easiest way to create and test a configuration:

pip install "cobair[web]"
cobair-web

It opens http://localhost:8000 in your browser, see the Web GUI documentation.

Qt Graphical User Interface

The Qt configurator is kept in maintenance mode (no new features). Install it with pip install "cobair[qt]" and start it with

cobair-qt

if you want to start the GUI with a loaded config use

cobair-qt -f config.yml

From a source checkout, python start_configurator.py [-f config.yml] does the same.

Tutorial

For a step-by-step guide on how to use CoBaIR, check out our Tutorial.

Documentation

The Documentation can be accessed on https://dfki-ric.github.io/CoBaIR/

Bayesian Approach

In the bayesian approach CoBaIR uses a two-layer Bayesian Net of the following structure.

Config Format

Configs will be saved in yml files. For convenience there is a configurator which can be started with

cobair-qt

Bayesian Approach

The configuration file for a two layer bayesian net for context based intention recognition follows the given format:

# List of contexts. Contexts can have different discrete instantiations.
# Number of instantiations must be larger than 1.
# For all discrete instantiations a prior probability must be given(sum for one context must be 1)
contexts:
  context 1:
    instantiation 1 : float
      .
    instantiation m_1 : float
  context n:
    instantiation 1 : float
      .
    instantiation m_n : float
# List of intentions. Intentions are always binary(either present or not)
# For every intention the context variables and their influence on the intention is given
# [very high, high, medium, low, very low, no] => [5, 4, 3, 2, 1, 0]
intentions:
  intention 1:
    context 1:
        instantiation 1: int # one out of [5, 4, 3, 2, 1, 0]
        .
        instantiation m_1: int # one out of [5, 4, 3, 2, 1, 0]
    context n:
        instantiation 1: int # one out of [5, 4, 3, 2, 1, 0]
        .
        instantiation m_n: int # one out of [5, 4, 3, 2, 1, 0]
  intention p:
    context 1:
        instantiation 1: int # one out of [5, 4, 3, 2, 1, 0]
        .
        instantiation m_1: int # one out of [5, 4, 3, 2, 1, 0]
    context n:
        instantiation 1: int # one out of [5, 4, 3, 2, 1, 0]
        .
        instantiation m_n: int # one out of [5, 4, 3, 2, 1, 0]
# decision_threshold is a float value between 0 and 1 which decides
# when an intention should be considered in inference.
# Probability must be greater than decision_threshold.
decision_threshold: float

How to contribute

If you find any Bugs or want to contribute/suggest a new feature you can create an Issue or a Pull Request or contact me directly via adrian.auer@dfki.de. See CONTRIBUTING.md for the development and versioning workflow.

Run tests

Tests are implemented with pytest. All dependencies are declared in pyproject.toml; the development tools are dependency groups (test, docs, lint, e2e), which need pip >= 25.1. Install the test dependencies with

pip install --group test -e ".[all]"

Then you can run

python -m pytest tests/

You can as well see the test report for a specific commit in the Tests workflow runs (artifact test-reports-py<version>)

Coverage

If you want to see coverage for the tests you can run

coverage run -m pytest tests/

Use

coverage report

or

coverage html

You can as well see the coverage for a specific run in the job summary of the Tests workflow runs

To show results of the coverage analysis.

Code quality

The Quality workflow runs on every Pull Request into main and on pushes to main and counts ruff linting issues and the docstring coverage measured by interrogate. Both tools are configured in pyproject.toml and pinned in its lint dependency group. Run them locally with

pip install --group lint
ruff check .
interrogate -c pyproject.toml .

The workflow works as a ratchet against .github/quality-baseline.json: it fails if the number of ruff issues goes up or the docstring coverage goes down. If the numbers improved, lower ruff_issues / raise doc_coverage in the baseline in the same Pull Request to lock the improvement in. The badges at the top of this README are only committed by the workflow on main; on other branches and Pull Requests the numbers are shown in the job summary.

The test coverage badge comes from the coverage job of the Tests workflow. It combines the coverage of the cobair package from the Python 3.12 tests with all extras and from the end-to-end tests, including the cobair-web server they start (configured in [tool.coverage.*] of pyproject.toml). It is only shown, not enforced. Measure it locally with

pip install -e ".[all]"
coverage run -m pytest tests/ && coverage combine && coverage report

Build docu

Documentation is implemented with the material theme for mkdocs.

Dependencies

Install all dependencies for building the docu with

pip install --group docs -e ".[infer,qt]"

Build

Build the docu with

mkdocs build

The documentation will be in the site folder.

Authors

Adrian Auer & Arunima Gopikrishnan

Citation

If you use CoBaIR please consider citating it.

@inproceedings{lubitz2023cobair,
  title={Cobair: A python library for context-based intention recognition in human-robot-interaction},
  author={Lubitz, Adrian and Gutzeit, Lisa and Kirchner, Frank},
  booktitle={2023 32nd IEEE International Conference on Robot and Human Interactive Communication (RO-MAN)},
  pages={2003--2009},
  url={https://ieeexplore.ieee.org/document/10309581/},
  doi={10.1109/RO-MAN57019.2023.10309581},
  year={2023},
  month=aug,
  organization={IEEE}
}

@inproceedings{pub12351,
  author = {Auer, Adrian and Arriaga, Octavio and Hassan, Teena and Hoyer, Nina and Kirchner, Elsa Andrea},
  title = {A Bayesian Approach to Context-based Recognition of Human Intention for Context-Adaptive Robot Assistance in Space Missions},
  booktitle = {Proceedings of SpaceCHI 2.0 - Human-Computer Interaction for Space Exploration - A Workshop at CHI 2022. SpaceCHI 2.0: Human-Computer Interaction for Space Exploration (SpaceCHI 2.0-2022), located at ACM CHI 2022, May 1-1, New Orleans, LA, United States},
  year = {2022},
  month = {5},
  publisher = {ACM}
}

Funding

CoBaIR is currently developed in the Robotics Group of the University of Bremen, together with the Robotics Innovation Center of the German Research Center for Artificial Intelligence (DFKI) in Bremen. CoBaIR has been funded by the German Federal Ministry for Economic Affairs and Energy and the German Aerospace Center (DLR). CoBaIR been used and/or developed in the KiMMI-SF project.