Physical AI · evidence and evaluation

Planning and feedback control do different jobs

A planner chooses what to do over a horizon; a feedback controller keeps the system on that choice despite disturbance. Learned policies can replace either, and confusing which one was replaced makes a result unreadable.

How it works

Planning reasons about goals, sequencing and feasibility, usually at tens or hundreds of milliseconds. Feedback runs far faster, with the job of rejecting disturbance and keeping contact forces and trajectories bounded. An end-to-end learned policy may be doing both, and that is where its evaluation gets difficult: a failure could be a bad plan, a bad correction, or a plan the controller could not execute. Maha’s proposal is to state, before evaluating, which layer the learned component replaces and which layer remains hand-designed, so failures can be attributed rather than argued.

A concrete case

A policy that pauses mid-reach when an object shifts may be re-planning, or may have saturated a velocity limit. Logging the commanded and achieved velocities separates those two explanations; a success-rate table does not.

What this establishes

Nothing empirical. This is a decomposition Maha proposes for reading results, consistent with ordinary control practice.

What it does not

Not a control-design guide, and no stability property is asserted for any learned controller.

Questions worth asking

  • Ask which layer is learned and which is classical, including any safety filter.
  • Ask at what rate each layer runs and what happens when the slower one is late.
  • Ask whether commanded and achieved quantities are logged separately.

Sources

This page proposes a method rather than reporting a finding about the world, so it cites no external source. Where it describes something published, that page carries the citation.

Continue

Elsewhere on this site

Maha Strategies publishes explanation and evaluation method. We build no robots, run no physical experiments, and report no benchmark results of our own. Hardware, safety and evidence-intake questions live in the robotics section.