Heterogeneous Federated Learning in Robotic Systems
Author
Przybyszewski, Jan Stefan
Term
4. semester
Education
Publication year
2022
Submitted on
2022-06-02
Pages
78
Abstract
I en datadrevet verden er privatliv en vigtig bekymring for brugere af robotsystemer. Federated learning (FL) er en maskinlæringsmetode, hvor flere klienter træner sammen uden at dele deres lokale data, hvilket hjælper med at beskytte privatlivet. I dette arbejde præsenteres et nyt, heterogent FL-rammeværk, der kan træne en føderation af klienter, selv når de bruger forskellige modelarkitekturer. Med dette rammeværk trænes modeller til at forudsige robotgreb, og en pick-and-place-arbejdsgang implementeres, hvilket præsenteres som den første anvendelse af FL i industrirobotik. Systemets fleksibilitet vises også på opgaver inden for billedklassifikation og sentimentanalyse. Derudover undersøges, hvordan vidensdestillation påvirker træningsresultaterne. Resultaterne viser, at rammeværket kan forbedre klienters ydelse markant sammenlignet med at træne i isolation. Samtidig reducerer det, i modsætning til standard distribuerede tilgange, de privatlivsrisici, der opstår ved at dele data.
In a data-driven world, privacy is a key concern for users of robotic systems. Federated learning (FL) is a machine learning approach where multiple clients train together without sharing their local data, which helps protect privacy. This work introduces a new, heterogeneous FL framework that can train a federation of clients even when they use different model architectures. Using this framework, the authors train robot grasp prediction models and deploy a pick-and-place pipeline, presented as the first application of FL in industrial robotics. The system’s flexibility is also demonstrated on image classification and sentiment analysis tasks, and the impact of knowledge distillation on training is investigated. Results show that the framework can significantly improve client performance compared to training in isolation. At the same time, unlike standard distributed approaches that involve data sharing, it reduces privacy risks.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Keywords
Documents