Federated Learning for Care Robotics

This repository contains an implementation of federated learning for a hierarchical approach to behavioral cloning in care robotics. By training locally and exchanging only model weights, we preserve user privacy and achieve competitive performance compared to a non-federated setup.

---

Table of Contents

1. Overview
2. Installation
3. Repository Structure
4. Usage
5. Technical Details
6. Further Reading

---

Overview

The project implements a three-tier hierarchy (H1, H2, H3) for robotic control:

• H1 (Classification #1)
• H2 (Classification #2)
• H3 (Regression)

These hierarchies build upon each other, which improves explainability by decomposing the learning process into interpretable stages. Meanwhile, federated learning ensures data stays local, mitigating privacy concerns.

---

Installation

Environment

1. Create a Conda environment with Python 3.10:

conda create -n fxai_env python=3.10 
conda activate fxai_env

2. Install Packages wit pip:

pip install -r requirements.txt

---

Data Preparation

1. Download the dataset from Google Drive: Videos_Database_20_Robot_WebCam_50_overall_database.npz

2. Create a folder named Data in the repository root (if it does not exist).

3. Place the downloaded .npz file inside Data/.

Repository Structure

Below is a simplified layout of important files:

. 
├── Data
│ └── Videos_Database_20_Robot_WebCam_50_overall_database.npz 
├── environment.yml 
├── README.md 
└── src 
  ├── config 
  │ └── config.yaml 
  ├── h1.py 
  ├── h2.py 
  ├── h3.py 
  ├── run_all.py 
  └── utils 
    ├── data_utils.py 
    ├── federated_utils.py 
    ├── logging_utils.py 
    ├── model_h1.py 
    ├── model_h2.py 
    ├── model_h3.py 
    └── training_utils.py

• h1.py, h2.py, h3.py: Each handles the federated learning process for its respective hierarchical model.
• run_all.py: Script that runs H1 → H2 → H3 in sequence, ensuring each stage starts from the model produced by the previous one.
• utils/: Contains supporting utilities for data loading, federated setup, logging, and model definitions.
• config/config.yaml: Hydra configuration specifying model parameters, data paths, etc.

---

Usage

1. Edit config/config.yaml to adjust hyperparameters, paths, or other settings.
2. Run the entire pipeline:

cd src 
python run_all.py

• A timestamped directory will be created under cfg.paths.output_dir, with subfolders for H1, H2, and H3.
• Logs, CSV metrics, and final .h5 models are saved accordingly.

3. Check logs in each subfolder to see training progress and final metrics.

To run only one hierarchy, you can execute python h1.py, python h2.py, or python h3.py individually, but note that h2.py and h3.py expect existing model weights from the previous stage.

---

Technical Details

• Federated Learning: Implemented via TensorFlow Federated (v0.58.0).
• Hierarchical Models:
• H1 learns a first-level classification.
• H2 refines or adds a second-level classification task.
• H3 focuses on regression using the final layer from H2.
• Each subsequent hierarchy loads and adapts the model from the previous stage.
• Explainability: By splitting the modeling task into multiple phases, the internal states (like “phase” or “state”) become more interpretable.

---

Further Reading

A broader discussion of Federated Learning in Care Robotics, including experimental results and a detailed presentation, is available in a separate repository: Federated-Learning-in-Care-Robotics

Feel free to explore that for in-depth information, additional data, and advanced experimentation details.