ACTSEL source code repository

We introduce ACTSEL. A method for automatic selection of actions that help optimally determine physical object properties that are not readily available through vision. Originally part of the publication: Interactive Learning of Physical Object Properties Through Robot Manipulation and Database of Object Measurements

Link to paper

Link to arXiv.

Graphical model overview

General overview of the algorithm (left), Bayesian network and action relations (right)

Running the model

For best experience install conda environment as numpy, scipy and scikit-learn are needed for algorithm operation

1. To run the model, fill in the templates for nodes, actions and their relevant confusion matrices in configs/templates. In order to update the actual config .json files, run the scripts/templates_to_cfgs.py from root directory as:

python3 scripts/templates_to_cfgs.py

2. Customize the main.py to meet your action and object requirements byt customizing experiment_object_names and action to node mapping.

Implementation remarks

The algorithm and results presented in the paper were obtained offline on pre-measured dataset for broader statistical understanding. This fact is reflected in main.py.

Citation

Please cite this paper as:

@article{kruzliak2024interactive,
      title={Interactive Learning of Physical Object Properties Through Robot Manipulation and Database of Object Measurements}, 
      author={Andrej Kruzliak and Jiri Hartvich and Shubhan P. Patni and Lukas Rustler and Jan Kristof Behrens and Fares J. Abu-Dakka and Krystian Mikolajczyk and Ville Kyrki and Matej Hoffmann},
      year={2024},
      eprint={2404.07344},
      archivePrefix={arXiv},
      primaryClass={cs.RO}
}