PolySLGen: Online Multimodal Speaking-Listening Reaction Generation in Polyadic Interaction

PolySLGen addresses a significant gap in embodied AI: the transition from dyadic (1-on-1) to polyadic (group) social interaction. While most current LLM-based agents struggle with group dynamics (who to look at, when to nod without interrupting, how to react to multiple speakers), this project provides a structured framework for 'online' (real-time) generation. Its defensibility is currently low (score 4) because, while technically complex, it is a research-grade implementation with 0 stars and 4 forks, indicating it has yet to build a community or integration ecosystem. The primary moat is the specialized logic for multi-user social signaling, which is more niche than general conversation. However, frontier labs like OpenAI (with GPT-4o's real-time capabilities) and Google (with Project Astra) are rapidly moving toward multimodal, low-latency social agents. The risk of platform domination is high because these companies control the underlying multimodal models and can easily integrate 'social group logic' as a system-level feature. PolySLGen is highly valuable as a reference for academic researchers or robotics startups (e.g., Figure, 1X) needing specific social interaction layers, but it faces rapid displacement if frontier models internalize multi-user turn-taking and non-verbal cues natively.