Swarm Descriptions: Creating and Demonstrating a Dataset for Swarm Mission Generation from Natural Language
For installation, the Hatch Project Manager is recommended.
Hatch can be installed with pipx install hatch, assuming pipx is installed. Pipx can be installed using most package managers or with python -m pip install --user pipx. Make sure to execute pipx ensurepath after installation.
From repository level start hatch virtual python environment with hatch shell and install python module with python -m pip install -e ..
Copy custom directory to /opt/argos/custom/. Execute xhost +local:docker to enable visualization.
Run scripts like python scripts/test.py -h inside the hatch virtual environment.
generate_data.pysamples descriptions and configurations for demonstration.parse_eval_datasets.pyconverts the dataset generated by the inference notebook for evaluation.eval_results.pycontains the data for evaluation of the finetuned model.run_config_params.pyruns configuration params xml generated from dataset (e.g.generate_data.py.)test.pyprints several properties generated from a randomly sampled mission.
Finetuning and inference was done on Kaggle.
figures.ipynbfigures and metrics for report.mistral-finetuning-swarm.ipynbfinetuning the LLM on our dataset. Includes relevant functions on how to generate dataset from our python module.mistral-inference-swarm.ipynbgenerating configuration params from descriptions by fined LLM for evaluation.
- Towards an integrated automatic design process for robot swarms by Bozhinoski et. al.
- Behavior Trees as a Control Architecture in the Automatic Modular Design of Robot Swarms by Kuckling et. al.
This work is part of my masters project at the University of Konstanz. Find the project report at project_report.pdf or the presentation at project_slides.pdf.
hatch shellcannot install wheel: python-version in pyproject.toml needs to be in the shape of ">=3.10" and not "==3.10.12". At least on Windows.