README.md

Contents

• Installation
  • Anaconda Environment Creation
  • Package Installation
• General Framework
• Environment
  • Download Collected Dataset
• Experiments
  • Dataset Analysis
  • Model Training
  • Model Evaluation
• Learning Trust
• Results
  • Experiment Outcomes
  • Optimized Model Parameters
• Corresponding Author

---

Installation

Anaconda Environment Creation

Requires Python >= 3.10:

• conda create -y -n trust_learning python=3.10
• conda activate trust_learning

Package Installation

Execute setup from the source code root of the repository:

• cd src
• pip install -e .

---

General Framework

Figure 1: Visualization of the Proposed Framework

---

Environment

A markov decision process (MDP) environment script is located in the following directory:

• cd src/environment

Figure 2: Environment of Tiling Operation with Cobot (@ UTS - Robotics Institute)

Download Collected Dataset

Human collection dataset of demonstrations is shared in the repository:

• cd dataset/human_demonstrations

Training and testing datasets of .json files are located in the following directory:

• cd dataset/human_demonstrations/2024_01_23_Train/jsons
• cd dataset/human_demonstrations/2024_02_02_Test/jsons

---

Experiments

Figure 3: Environment of Human-Robot Teaming in Construction (@ UTS - Robotics Institute)

Figure 4: Human Demonstration Data Collection Process

Dataset Analysis

Priorly collected dataset is shared in the repository:

• cd dataset/human_demonstrations

Figure 5: Human Collected Training Dataset of 43 Trajectories (Expert Demonstrations)

The dataset is analyzed in the following notebook:

• cd src/analyses
• jupyter notebook
• visualize_demonstration.ipynb

Figure 6: Visualization of One Trajectory of Human Operation (Expert Demonstration)

Model Training

Neural network model files are located in the following directory:

• cd src/models

To run the training script:

• cd src/optimization
• python train.py

To run the training of the policy model (explicitly):

• cd src/optimization
• python train_policy.py

To run the training of the reward model (explicitly):

• cd src/optimization
• python train_reward.py

Model Evaluation

The evaluation of the trained models is included in the following notebook:

• cd src/evaluation
• jupyter notebook
• evaluate_trust_estimation.ipynb

Figure 7: Parameter Optimization during Robot Learning of Human Trust after Each Experiment

Figure 8: Relationship Between Human Trust Measurements and Learned Reward Function

Figure 9: Shifts in Trust Distribution During Human Trust Learning Stage Experiments

A Beta Reputation System implementation script is located in the following directory:

• cd src/trusts
• model_dynamics.py

---

Learning Trust

Figure 10: Illustration of an Iterative Trust Estimation Process

---

Results

Visualize the results of the Modeled Human Trust in the following notebook:

• cd src/evaluation
• jupyter notebook
• visualize_trust_dynamics.ipynb

Figure 11: Comparing Trust Estimation with Trust Measurement in Verification Stage Experiments

Figure 12: Reward Function and Probabilistic Trust Estimation in the Verification Stage Experiments

Experiment Outcomes

The resultant Excel files of the experiments are shared in the following directory:

• cd results/experiments

Trust learning stage experiment results:

• cd results/experiments/learning_experiments
• results/experiments/learning_stage_experiment_results.xlsx

Inference stage experiment results:

• cd results/experiments/inference_experiments
• results/experiments/inference_stage_experiment_results.xlsx

Optimized Model Parameters

The trained policy network and reward model parameters are shared in the following directories:

• cd results/policy_network_params
• cd results/reward_network_params

---

Corresponding Author

For any inquiries or lack of clarity, please contact the corresponding author: ➔ [email protected]