Robot Agentic OS / 具身智能体操作系统

Definition

A robot Agentic OS is an embodied agent stack that composes speech recognition, intent understanding, task decomposition, task splitting, tool invocation, and low-level robot interfaces into one operating system, with System 1 execution and System 2 planning coordinated by a harness.

Current Synthesis

The source says a robot must be an agent system that interacts with people and the environment, not a single point model, and that the company started building an agentic system early. Its described stack includes speech recognition, task decomposition, task splitting, tool calling, and low-level interfaces. System 2 understands intent and plans; System 1 executes quickly and finely. The upper model was initially an external ByteDance Seed model and the company is experimenting with replacing it with its own, while the lower action layer accumulates as the company’s own asset. The harness matters because it manages context and makes the brain naturally schedule a stable, controllable small-brain model; the company treats that coordination as design work that model capability alone does not supply.

Key Claims

• A robot should be treated as an agent system rather than a single model.
• The described stack includes speech recognition, task decomposition, task splitting, tool calls, and low-level interfaces.
• System 2 handles intent understanding and planning while System 1 handles fast, fine execution.
• The upper layer can be borrowed or swapped from an external provider, while the lower action layer is the accumulating company asset.
• The harness, meaning context management, brain/small-brain interaction, and scheduling, is a distinct layer of engineering.
• Startup compute constraints shape the build order: the source says priority may sit with Action pretraining rather than the upper model.

Evidence

• Agent-system framing: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 says a robot is necessarily an agent that interacts with people and the environment, not a single model.
• Stack evidence: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 lists speech recognition, task decomposition, task splitting, tool invocation, and low-level interfaces, and describes the early Agentic System build.
• System 1/System 2 evidence: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 assigns intent understanding and planning to System 2 and fast fine execution to System 1, with an external upper model at the start.
• Harness evidence: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 says the harness handles context management and lets the brain schedule a stable, controllable small-brain model, and that these coordination designs do not follow automatically from model capability.
• Compute-priority evidence: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 says startup compute is precious and may go to Action pretraining because upper foundation models are still changing quickly.

Counterevidence & Qualifications

The description is a company’s own account from one interview and does not include interfaces, latency, autonomy boundaries, or failure handling. The source also leaves open whether the upper layer eventually absorbs System 1, so the architecture should be read as a current build choice under fast-moving external models rather than as a settled endpoint.

What Changed

• Created the concept from the interview’s description of an embodied Agentic OS with System 1/System 2 and a harness.
• Recorded the swap-prone upper model and the accumulating lower action layer as the startup’s risk-hedging structure.

Related Concepts

• Agent Harness - general harness layer that the robot OS specializes for physical action.
• Small Brain Action Layer / 小脑（Action Policy） - System 1 execution layer the OS has to schedule.
• Embodied Context & Memory / 具身上下文与记忆 - context and memory that the harness manages.
• AI Model Orchestration - orchestration pattern that the robot OS applies to models and tools.
• General Model Robot Boundary - boundary the OS bridges between a general upper model and physical execution.