Federated Continual Learning for Socially Aware Robotics

Luke Guerdan,Hatice Gunes

2023-07-10

Abstract:From learning assistance to companionship, social robots promise to enhance many aspects of daily life. However, social robots have not seen widespread adoption, in part because (1) they do not adapt their behavior to new users, and (2) they do not provide sufficient privacy protections. Centralized learning, whereby robots develop skills by gathering data on a server, contributes to these limitations by preventing online learning of new experiences and requiring storage of privacy-sensitive data. In this work, we propose a decentralized learning alternative that improves the privacy and personalization of social robots. We combine two machine learning approaches, Federated Learning and Continual Learning, to capture interaction dynamics distributed physically across robots and temporally across repeated robot encounters. We define a set of criteria that should be balanced in decentralized robot learning scenarios. We also develop a new algorithm -- Elastic Transfer -- that leverages importance-based regularization to preserve relevant parameters across robots and interactions with multiple humans. We show that decentralized learning is a viable alternative to centralized learning in a proof-of-concept Socially-Aware Navigation domain, and demonstrate how Elastic Transfer improves several of the proposed criteria.

Machine Learning,Human-Computer Interaction

What problem does this paper attempt to address?

### Problems Addressed by the Paper

This paper aims to address two major challenges faced by social robots in practical applications:

1. **Insufficient Behavioral Adaptability**: Current social robots are designed for single interactions and rely on pre-configured capabilities, whereas future systems need to adapt over multiple long-term interactions.
2. **Insufficient Privacy Protection**: Existing centralized machine learning methods (which centralize training data on a single machine or server) lead to privacy issues, limiting user trust and the widespread application of social robots.

To solve these two problems, the authors propose a decentralized learning paradigm—Federated Continual Learning for Socially Aware Robotics (FCL4SR), which combines Federated Learning (FL) and Continual Learning (CL) methods. Specifically:

- **Federated Learning**: Collaborates between distributed devices (such as robots) to exchange model information instead of storing raw data on a central server.
- **Continual Learning**: Enables robots to continuously learn new intelligent behaviors over multiple interactions without forgetting previously acquired knowledge.

In this way, FCL4SR not only enhances the personalization capabilities of social robots but also achieves these functions while protecting user privacy.

### Main Contributions

1. **Proposing the Decentralized Learning Problem for Social Robots**: For the first time, the social robot problem is formulated as a decentralized learning problem combining FL and CL.
2. **Developing a Decentralized Learning Evaluation Framework**: A set of evaluation criteria is proposed, including adaptation quality, adaptation time, knowledge sharing, and model overhead, and a comprehensive evaluation of existing methods is conducted.
3. **Proposing New Strategies to Improve FCL Performance**: A new regularization algorithm—Elastic Transfer—is proposed to improve knowledge sharing, adaptation quality, and adaptation time in FCL methods.