估值超170亿元，头部具身智能大模型创企发布最新VLA模型！家庭服务机器人，要来了！

Core Insights - The article discusses the launch of the VLA model π0.5 by Physical Intelligence (PI), a startup valued at over $17 billion, which showcases advanced capabilities in performing complex household tasks in unfamiliar environments [1][16]. Summary by Sections Introduction of π0.5 Model - The π0.5 model is an advanced visual-language-action (VLA) model that allows robots to perform long-duration, complex household tasks such as cleaning kitchens and organizing bedrooms, demonstrating superior generalization capabilities in open-world scenarios [1][2]. Functionality and Training - The π0.5 model emphasizes functional transfer in scenarios not covered by training data, relying on both physical manipulation skills and an understanding of environmental "common sense," which includes object recognition and semantic reasoning [2][5]. - The model learns from heterogeneous data sources, enabling it to understand the semantic context of tasks and break down task steps effectively [5][8]. Architecture and Decision-Making - The architecture of π0.5 employs a dual-system framework, integrating high-level decision-making and low-level execution within the same model, allowing for a cohesive approach to task execution [8][10]. - The model generates high-level action plans expressed in language and matches them with motion instructions, continuing the development of the Hi Robot system [10][11]. Industry Context and Competitors - Since 2025, the dual-system architecture has become mainstream in the field of embodied intelligence, with leading companies like Figure AI and Nvidia also adopting similar models [14]. - The article highlights the competitive landscape, noting that several companies, including domestic players like Zhi Ping Fang and Ling Chu Intelligent, are developing their own VLA models with dual-system architectures [14][21]. Challenges and Future Directions - While π0.5 shows promise, it still faces challenges in high-level semantic reasoning and action execution, indicating that further advancements are needed to achieve flexible physical intelligence [13][15]. - The article suggests that the integration of large models and advanced algorithms will be crucial for the commercialization and functionality enhancement of humanoid robots [20][21].