Something changed in how I look at @axisrobotics after seeing the latest research update.
The interesting part is not simply that the dataset keeps getting bigger.
The Axis research has now grown to more than 1,800 tasks and 1.5M+ trajectories, and the work has been accepted to the Physical World Models workshop at IROS 2026.
That tells me something important.
Axis is starting to look less like a platform that simply collects robot data and more like an infrastructure for continuously studying how robot learning scales.
More tasks create more behavioral coverage.
More trajectories expose more edge cases.
Those results can then be used to understand where models still struggle and what kind of data should come next.
That is a very different way of thinking about scale.
You are not just asking:
“How much data do we have?”
You are asking:
“What did the last round of data teach us, and what should we collect next?”
For Physical AI, I think that question matters just as much as the raw dataset size.
The real advantage may come from the system getting better at learning from its own data as it grows.