Layered Robot Architecture

Definition

Layered robot architecture separates slower high-level understanding, planning, and supervision from faster low-level perception-action and control, while allowing training or deployed behavior to remain learned and partly end-to-end.

Current Synthesis

The complete source set converges on layering for different reasons. The Sudu interview argues that scarce real-robot data and weak open-world manipulation favor an upper planner plus reliable lower-level skills trained with structured 3D information and simulation. The Physical Intelligence interview keeps learned policies and real-machine experience central, but shows that traditional planning, control, and performance discipline remain part of modern robot learning.

The Q3 review makes the upper layer more general. Astra can reason about changed scenes, generate end-effector targets, supervise a specialized policy, and call simulation tools, while a local action model handles speed, precision, contact, and continuous feedback. Layering is therefore not a retreat to brittle hand-coded pipelines; it is a way to allocate different timescales, data types, and reliability requirements to the components best able to meet them.

Key Claims

• The upper layer interprets the environment, decomposes tasks, chooses strategies, and can supervise or correct execution.
• The lower layer performs fast, contact-rich skills such as grasping, insertion, pouring, turning, and disturbance recovery.
• Layering can use learned geometry, material, kinematic, future-state, and action representations rather than fixed rules alone.
• Reliable short skills remain prerequisites for long tasks because serial failures compound.
• A general foundation model can deepen the upper layer without eliminating embodiment-specific policies, controllers, sensors, or hardware validation.
• Deployment may appear end-to-end even when pretraining and system design expose interpretable intermediate structure.

Evidence

Structured manipulation and simulation

• E244｜端到端vs上下分层：机器人路径之争，正在转向？ defines the upper planning layer and lower manipulation layer and argues that structured 3D data, Sim2Real, and hardware/software co-design can improve open-world reliability.

Learned-policy continuity

• 146. 对Physical Intelligence柯丽一鸣4小时访谈：Pi的开源模型研究，机器人的江湖、族谱与主角 connects contemporary imitation, reinforcement learning, and VLA policies to traditional robotics concerns about planning, control, and real-machine performance.

General-model upper layer

• 184: 与陈哲的「具身季报26Q3」：Astra冲击后，通用模型会吃掉Robotics吗？ reports Astra generating end-effector targets and supervising another action model while specialized local policies retain an advantage in precise, high-frequency contact tasks.

Counterevidence & Qualifications

The sources do not establish one universally optimal boundary between layers. Extra interfaces can introduce latency, error propagation, and integration complexity, while end-to-end training may discover useful representations that designers omit. The reported Astra success figures lack common protocols and remain far below production proof. Hardware, task, safety, and available data can shift which functions belong in each layer.

What Changed

• Migrated the page to the synthesis-v1 schema while preserving its complete evidence inventory.
• Added a general foundation model as planner, supervisor, and tool user above a specialized action layer.
• Reframed the architecture around timescale and reliability allocation rather than a simple end-to-end-versus-modular split.

Related Concepts

• General Model Robot Boundary - determines which responsibilities can move into a general upper layer.
• Small Brain Action Layer / 小脑（Action Policy） - local real-time execution layer beneath slower reasoning.
• Vision Language Action Models - learned policy family that can occupy or bridge layers.
• Open-World Robot Manipulation - reliability bar for lower-level skills outside rehearsed scenes.
• Structured 3D Robot Data - intermediate physical representation used by one layered route.
• Robot Response Latency - timing constraint that motivates separating slow reasoning from fast action.
• Physical Recursive Self-Improvement / Physical RSI - tool-using simulation loop that may improve policies across the architecture.