Why household-chore data does not transfer to skilled manipulation tasks

The dominant source of human manipulation data in 2025 is still crowd-sourced household footage: someone picks up a cup, stacks dishes, folds laundry. It is cheap to collect, it scales, and it looks plausible. It also transfers poorly to professional trade tasks—and the reasons are instructive.

The distribution problem

Household tasks occupy a narrow slice of the manipulation space. Objects are light and forgiving; failure is recoverable; the task environment is largely unstructured. Skilled trades work lives in a different part of the distribution entirely. A licensed electrician routing cable through a conduit is applying calibrated torque, maintaining precise tool alignment under load, and making judgment calls about material state that are invisible to a layperson watching the footage.

The gap is not just about domain labels. It is a gap in the underlying physics regime. Household manipulations are predominantly in the low-force, slow-speed, unconstrained contact regime. Electrical and plumbing work involves constrained contact, high force-to-weight ratios, and tools that amplify small positional errors into large outcome failures. A model trained primarily on the first distribution will not generalize to the second, because the prediction targets are structurally different.

Why scale does not fix it

The intuition that "more data fixes distribution shift" has a limit. It applies when the target distribution is a harder version of the same distribution, not when it is a different distribution. Adding ten million more household episodes does not add a single example of a cable gland being torqued to a specified value, because that action does not appear in household footage.

The analogous failure mode in vision was training general-purpose classifiers on ImageNet and expecting them to classify microscopy slides. More ImageNet did not help—different data was required.

What the distribution actually requires

Training on professional trade demonstrations requires footage that captures the right force regime, the right tool-object contact geometries, and critically, the failure and recovery modes that define competence in the domain.

A journeyman electrician makes mistakes and corrects them. That correction sequence—the verify span that returns a failure result, followed by the recover subtask that brings the attempt back to a passing state—is the most expensive demonstration to collect and the most valuable for learning. It does not appear in household footage because household tasks do not have formal success criteria.

The provenance problem

There is a second-order issue that household footage compounds. Crowd-sourced video has opaque provenance: unknown capture conditions, unknown demonstrator skill level, unknown task completion criteria. A model trained on this footage cannot be interrogated—you cannot pull the episodes where a particular subtask failed, because the footage was never annotated at that level.

Professional demonstrations can be structured with explicit episode boundaries, subtask annotations, outcome labels, and per-episode provenance records. That structure is what turns footage into a dataset a training team can actually reason about: you can filter by outcome, you can isolate failure modes, you can check that your eval split contains demonstrators your training split never saw.

The data problem in physical AI is not just volume. It is distribution, structure, and provenance—and those three things require a different collection approach from the start.