Year: 2025
Type: dissertation
Abstract: This thesis explores how a social robot can autonomously generate adaptive non-verbal sounds to enhance its expressiveness and communicative capacity for Human-Robot Interaction (HRI), utilizing Blossom, an open-source platform. The key challenge addressed is the limited adaptability of current social robots, which often rely on static, predefined sound responses. This limits their ability to respond dynamically to varying interaction contexts. While significant progress has been made in speech-based interaction, non-verbal sound remains underexplored, particularly as a modality for conveying emotional or situational information in a language-independent way. To address the bridge, we propose a learning-based sound generation framework based on reinforcement learning with human-in-the-loop feedback. A Proximal Policy Optimization (PPO) agent learns to generate expressive sound parameters by interacting with human. To speed up learning, we introduce a warm-up phase based on supervised pretraining from labeled sound data. A hierarchical label classifier is also used to generate intent tags that condition the agent’s behavior. Experiments show that the pretraining phase improves initial policy performance and stabilizes early learning. The RL agent, when fine-tuned with human feedback, demonstrates improved ability to generate contextually and emotionally appropriate sounds. Compared to an untrained policy, the learned agent achieves higher reward scores and greater consistency under human evaluation. Overall, this work presents a practical approach for enabling social robots to produce adaptive non-verbal sounds. It contributes to closing a gap in the field and lays the groundwork for future research into more complex robots’ sound expressiveness.
Author: Luo, Qiaolan
Language: English
Subfield: Social Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W6982303577
