Artificial Trust-Based Task Allocation (ATTA)

Dataset and software for the paper "Heterogeneous human–robot task allocation based on artificial trust. Sci Rep 12, 15304." (Ali, A., Azevedo-Sa, H., Tilbury, D.M., & Robert Jr., L.P., 2022). DOI: https://doi.org/10.1038/s41598-022-19140-5.

Dataset

There are 10 sets of N=500 tasks for allocation found in code/results/tasks.

Software

Dependencies

All implementations were tested with Python 3.8.3 and PyTorch v1.7.1. The following packages are needed (please install with python3 -m pip install --user <package name>):

• numpy
• torch
• pickle
• scipy
• sklearn
• random
• sys
• matplotlib
• math
• time
• csv
• pandas

Task Allocation Methods Implementation

The ATTA method and code follows these steps in code/atta_caseI.py and code/atta_caseII.py:

• fabricate capabilities for both the human and robot
• load a .mat file with N tasks for allocation
• for each task:
  • compute the robot's trust in the human and robot's self-trust
  • compute the robot's and human's expected total reward
  • allocate the task to either the human or the robot
  • observe the outcome of the task as either a success or a failure
  • if the task was allocated to the human, update the human's capabilities belief distribution lower and upper bounds. This step is only done in code/atta_caseII.py.

There are two other task allocation methods evaluated in the paper. The random task allocation method and code follows these steps in code/random_caseI.py:

• fabricate capabilities for both the human and robot
• load a .mat file with N tasks for allocation
• for each task:
  • randomly allocate the task to either the human or the robot
  • observe the outcome of the task as either a success or a failure

Tsarouchi et al.'s task allocation method was implemented following the paper "On a human-robot collaboration in an assembly cell. International Journal of Computer Integrated Manufacturing, 30(6), 580-589." (Tsarouchi, P., Matthaiakis, A. S., Makris, S., & Chryssolouris, G., 2017). DOI: https://doi.org/10.1080/0951192X.2016.1187297. The implementation of Tsarouchi et al.'s task allocation method and code follows these steps in code/tsarouchi_caseI.py and code/tsarouchi_caseII.py:

• fabricate capabilities for both the human and robot
• load a .mat file with N tasks for allocation
• for each task:
  • if only one agent is capable of the task, allocate the task to that agent
  • if neither agent is capable of the task, discard the task
  • if both agents are capable of the task, allocate the task to the agent with the lower cost
  • observe the outcome of the task as either a success or a failure

Use Instructions

Task Allocation for Case I and Case II

1. Run python generate_normaldist_tasks.py from the code directory. This will generate .mat files with tasks for allocation in code/results/tasks.

2. Run the appropriate task allocation method(s) from the code directory. This will generate .mat files with the results of the task allocation in the respective directories of code/results/atta, code/results/random, and code/results/tsarouchi. The choices are:

• python atta_caseI.py
• python atta_caseII.py
• python random_caseI.py
• python tsarouchi_caseI.py
• python tsarouchi_caseII.py

3. Generate .csv files for data analysis. This will prepare the data for statistical testing and analysis.

• For case I, run python significance_prep_caseI.py which will generate code/results/spss/dataprep_caseI.csv and run python other_data_caseI.py which will generate code/results/spss/other_data_caseI.csv.
• For case II, run python significance_prep_caseII.py which will generate code/results/spss/dataprep_caseII_pos10.csv and code/results/spss/dataprep_caseII_neg10.csv and run python other_data_caseII.py which will generate code/results/spss/other_data_caseII_pos10.csv and code/results/spss/other_data_caseII_neg10.csv.
• For ATTA in case II only, run python explore_atta_results.py which will generate code/results/spss/dataprep_attaonly_caseII.csv.

4. Explore the results and data using IBM SPSS 26 and MATLAB.

Data and Results

The results from each task allocation method for 10 simulations are in the respective directories of code/results/atta, code/results/random, and code/results/tsarouchi.

The results in code/results/spss are from significance testing using IBM SPSS 26.

The results can be further explored in MATLAB using the scripts in code/results/matlab_scripts.

Paper Source Files and Figures

The paper directory also contains the LaTeX source files for the paper. Paper figures are in .svg format in paper/SVG figures and in .pdf format in paper/PDF figures.

Misc

There are other README files in various directories that provide further information on how to use the code.