Tech notes.
Measurement problems we hit building the corpus, written up rather than kept internal. No results are withheld because they are unflattering.
Five supply models, not one market
Why the cheapest hour is free while demand grows, how supply split into five structurally different models, and what became scarce instead.
ReadHuman2Robot: the embodiment gap is partly a representation problem
A human hand trajectory is not a robot action. What transfers is contact, object motion, constraints and outcome.
ReadWhy these four thresholds
Deriving ingest gates from the physics of head-mounted capture — and naming the one that is a convention.
ReadCan robots learn from YouTube? A pipeline, and the arithmetic behind it
From rights gating through candidate discovery, 3D, contact and state transitions to training routing.
ReadThe negative space of retargeting
What physics replay throws away, and why it is the densest signal in the pipeline.
ReadThe physical data engine: building the data layer for generalist robots
A dataset ages. An engine that knows what it holds, what it lacks and what to collect next does not.
ReadThe physical data scaling law: what to scale, and where the curve bends
Scale still matters. What needs scaling is the state space, not the hour count.
ReadRobot-ready is not a label. It is a measurable readiness profile.
Conventional data QA proves a media file is sound. It says nothing about whether the data can train a robot.
ReadThe occlusion problem: why ego video alone is not enough
The most important interactions happen where the camera cannot see them. Occlusion is a structural property of embodied data.
ReadOne moment, six representations: the case for frame-level alignment
High-density data is not more sensors. It is the same physical event represented in several ways that can supervise each other.
ReadThe yield problem: why so much physical AI data never reaches training
The value in a data program sits in end-to-end yield, not in recorded hours.
ReadContact is the hidden token of physical intelligence
Language models learn structure from tokens. One of the decisive discrete events in physical AI is contact.
Read