对话星炽动力CTO李达——PULSE架构：如何让具身智能更像人？
Merged summary
TL;DR - Star Dynamics CTO Li Da introduces PULSE, a home-robot architecture that jointly models physical environments and users’ intentions, beliefs, and preferences. It aims to make robots safer, more personalized, and privacy-preserving in low-tolerance household settings.
- PULSE’s dual-track world model cross-validates environmental dynamics with predicted user states before acting.
- Its three-tier memory combines structured rules, RAG-based episodic retrieval, and foundation-model reasoning.
- Cloud training supports global improvement, while on-device test-time adaptation learns household-specific behavior from feedback and out-of-distribution data.
- Privacy-by-design keeps sensitive perception local and sends only abstractions such as poses, semantic labels, and scene graphs to the cloud.
Sources (1)
对话星炽动力CTO李达——PULSE架构：如何让具身智能更像人？
TL;DR - Star Dynamics CTO Li Da introduces PULSE, a home-robot architecture that jointly models physical environments and users’ intentions, beliefs, and preferences. It aims to make robots safer, more personalized, and privacy-preserving in low-tolerance household settings.
- PULSE’s dual-track world model cross-validates environmental dynamics with predicted user states before acting.
- Its three-tier memory combines structured rules, RAG-based episodic retrieval, and foundation-model reasoning.
- Cloud training supports global improvement, while on-device test-time adaptation learns household-specific behavior from feedback and out-of-distribution data.
- Privacy-by-design keeps sensitive perception local and sends only abstractions such as poses, semantic labels, and scene graphs to the cloud.