July 21, 2026 · Tuesday2026 年 7 月 21 日 · No. 17 · updated更新于 07:35 () SIGNALS信号EVENTS日历ARCHIVE归档ACCESS接入ABOUT关于

AUTOSIGNAL

Human-curated. Expert-annotated. Every signal traced to source. 人工精选 · 专家点评 · 每条信号可溯源

← All signals← 返回全部信号

Research & IP研究与专利

PhyAgentOS: Runtime System Unifies Scheduling, Verification, and Memory for Embodied AgentsPhyAgentOS:为具身智能体统一调度、验证与记忆的运行时系统

S 1.7 T1 1 sources1 个来源 R7-research
  1. Researchers led by Yang Liu (11 authors total, arXiv:2607.16636, submitted July 18, 2026) present PhyAgentOS, a runtime foundation that delivers scheduling, verification, memory, benchmarking, and safety as system-level services for embodied agents.
  2. The paper argues that vision-language-action models, world models, and agentic planners each advance physical intelligence individually, but their combination currently lacks a common execution abstraction, shared state, semantic verification, and persistent experience across heterogeneous robot embodiments.
  3. PhyAgentOS's core mechanism, the Session-Centered Runtime, treats an entire session—not a single action—as the minimum unit of scheduling, an architectural choice with direct relevance to automotive AI stacks that must likewise coordinate perception, planning, and safety verification across multi-step autonomous-driving tasks.
  1. Yang Liu 等 11 位作者(arXiv:2607.16636,2026年7月18日提交)提出 PhyAgentOS,一个为具身智能体提供调度、验证、内存、基准测试和安全等系统级服务的运行时基础设施。
  2. 论文指出,视觉-语言-行动模型、世界模型和代理式规划器虽各自推动了物理智能的发展,但三者组合目前缺乏统一的执行抽象、共享状态、语义验证以及跨异构机器人本体的持久化经验积累。
  3. PhyAgentOS 的核心机制“会话中心运行时”将整个会话(而非单一动作)作为调度的最小单元,这一架构思路与需要在多步自动驾驶任务中协调感知、规划与安全验证的汽车 AI 技术栈直接相关。