Small Brain Action Layer / 小脑（Action Policy）

Definition

The action policy, or “small brain,” is the lower layer of an embodied system that turns high-level intent into fast, stable, controllable physical action in real time, as distinct from the upper brain that reasons, plans, and interprets the world.

Current Synthesis

The source’s central architecture claim is that a general model is not enough for a robot. ChatGPT 6 / Astra driving a robot arm impressed on grasping and on multi-step spatial and task planning, including a pen drawing of the Golden Gate Bridge that improved over repeated attempts. But the source notes that demonstration videos are often sped up, that a single model inference takes a long time, that success in a static scene does not transfer to a dynamic one, and that a tilted cup may not be recovered fast enough. Those gaps define the small brain: a layer that can react, adjust, and keep the body controllable while the upper brain plans. The same split appears in the company’s System 2 planner and System 1 executor, and it explains why a compute-constrained startup may prioritize Action pretraining over the upper model.

Key Claims

• Grasping and task-planning success does not establish real-time physical control, because recovery from disturbance is a different capability.
• Sped-up demonstration video can hide inference latency, and single-pass inference time is a first-order constraint for physical tasks.
• The small-brain layer exists to provide fast, stable, controllable reactions that a reasoning model cannot guarantee.
• The brain/small-brain division maps onto System 2 intent understanding and planning versus System 1 fast fine execution, mediated by a harness.
• A startup may allocate scarce compute to Action pretraining while borrowing a fast-improving upper foundation model.
• The source thinks the lower layer eating the upper layer is unlikely, while the upper layer absorbing System 1 is possible with uncertain timing; a unified very-large embodied model is theoretically imaginable but cannot skip the intermediate system work.

Evidence

• Astra boundary 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 describes the robot-arm demonstration, the sped-up video, the long single-pass inference, and the static-versus-dynamic gap.
• Disturbance 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 uses the tilted cup as the case where a general model may not recover quickly enough without a reactive lower layer.
• Architecture 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 names the Action Policy or “small brain” as the lower layer and System 1/System 2 as the company’s operating split.
• 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 scarce and may go to Action pretraining because upper-layer foundation models are still moving fast.
• Convergence 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 gives the “below cannot eat above, above may eat System 1” judgment and the 10T unified-model thought experiment.

Counterevidence & Qualifications

The claim that the upper layer could eventually absorb System 1 is the source’s own caveat against a hard architectural separation, and the source gives no timing. The page also lacks model architecture, latency numbers, and success rates, so it records an argument structure rather than a measured comparison, and part of the demonstration evidence comes from public video that the source itself says is sped up.

What Changed

• Created the concept to hold the source’s brain/small-brain distinction separately from general layered-robotics discussion.
• Added the real-time disturbance-recovery boundary as the reason a general model cannot be the whole embodied system.

Related Concepts

• Layered Robot Architecture - broader upper-planning and lower-manipulation split into which this concept fits.
• Robot Response Latency - latency constraint that makes a reactive lower layer necessary.
• General Model Robot Boundary - boundary that this concept draws at real-time recovery rather than at semantics.
• Robot Agentic OS / 具身智能体操作系统 - system layer that schedules the small brain and mediates the brain-to-body interaction.
• Vision Language Action Models - model family that supplies action prediction but does not by itself settle the real-time control question.