Podcast Atlas

Updated · 2 episodes · 2 shows · 2 source notes

concept Topics: Technology

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 bounded sources agree that a general model is not enough for every robot-control cycle. ChatGPT 6 / Astra demonstrates grasping and multi-step planning, but sped-up video, slow inference, static-scene success, and disturbance recovery expose a separate latency and controllability problem. The small brain is therefore a body-specific layer that reacts, adjusts, and keeps execution stable while an upper model handles semantics, planning, review, and difficult cases. The newer source makes the operating split more concrete: a modest local control model can execute most actions, escalating likely failures to a stronger model and then a human, while recovery traces improve the local policy.

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 provides fast, stable, controllable reactions that a reasoning model cannot guarantee and should be co-designed with the robot body’s joints and control parameters.
  • 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 newer source suggests roughly 7B-to-16B local control models as a source-scoped design range.
  • The lower layer can learn from selectively captured failure-recovery trajectories, while eventual upper-model absorption of System 1 remains possible but unproven and does not remove today’s integration work.

Evidence

Counterevidence & Qualifications

The claim that the upper layer could eventually absorb System 1 is a caveat against hard architectural separation, and neither source gives a timing or controlled comparison. The page lacks disclosed model architecture, latency, success, failure-detection, or intervention-rate measurements; the 7B-to-16B range is a practitioner’s recommendation, not a validated universal optimum. Part of the demonstration evidence also comes from public video that the earlier source says is sped up.

What Changed

  • Added body-specific policy co-design and a source-scoped 7B-to-16B local-model range.
  • Extended the architecture from a static brain/small-brain split to failure-triggered escalation and recovery learning.
  • Narrowed the evidence boundary around failure detection, intervention rates, and model-size guidance.

Sources

2 source notes across 2 shows
  1. 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 十字路口Crossing
  2. 少有的同时做过 Agent、多模态和具身的人｜对谈前混元多模态 RL Infra 负责人淏楠 42章经