Physical AI · evidence and evaluation

The loop is the system: perception, state, action, consequence

A physical AI system is a loop, not a pipeline. Its own actions change what it will observe next, which is why accuracy measured on a fixed dataset does not predict behaviour when the policy is driving.

How it works

Offline, a perception model sees a dataset someone else collected. Online, it sees the consequences of its own choices: a slightly wrong grasp moves the object, the next image is one that no dataset contained, and the error compounds. Ross and colleagues made exactly this point for imitation learning, targeting performance under the distribution of observations the learned policy induces rather than the expert’s. The same structure governs latency: a decision computed from a 150 ms-old observation is applied to a world that has moved, so the effective loop delay belongs in the system description alongside the model’s accuracy.

A concrete case

The runnable fixture in this section makes this visible. A tracker with small per-step error and a two-step actuation delay drifts steadily away from its target, while the same tracker with zero delay converges — same perception quality, different closed-loop outcome.

What this establishes

That the induced-distribution problem is a recognised property of policies trained on someone else’s trajectories, stated as such in the imitation-learning literature.

What it does not

Recognising the problem does not quantify it for a given system. The fixture is a simulation and says nothing about any real robot’s stability.

Questions worth asking

  • Ask whether the reported accuracy came from a held-out dataset or from running the policy in the loop.
  • Ask for the end-to-end latency from exposure to actuation, not the inference time alone.
  • Ask what the system does when an observation is late or missing.

Sources

  • Ross, Gordon and Bagnell — A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning (DAgger) ↗

    A policy trained on an expert’s demonstrations encounters, at execution time, the distribution of observations that its own actions induce; the paper’s iterative approach targets good performance under that induced distribution rather than under the expert’s.

    Boundary: A learning-theoretic result with its own assumptions and benchmark experiments. It does not establish that any particular robot policy is safe, nor how much data a given task needs.

    Locator, anchor and reuse basis

    Read at: Abstract; §1 Introduction. Inspected 2026-09-19. Original paraphrase and link to the open preprint; no figures, tables or text reproduced.

    Verify by searching the source for: good performance under the distribution of observations it induces. If that phrase is not there, or does not carry the meaning stated above, this citation is wrong and we want to know.

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.