What if creating a dataset was the fun part?

Also published on LinkedIn.

In my last article I wrote about the dataset wall — the tedious, fragile grind of recording hundreds of teleoperated episodes that stops most newcomers before they start.

Here’s the thing: most of that grind is repetition, not learning. Demonstrating the same pick-and-place for the 60th time teaches the human nothing. So we asked: what if the robot did the repeating, and the human did the demonstrating?

That’s the core of how Raven creates datasets: kinematic action replay.

You demonstrate the motion once — guide the arm through the task the way you’d show a person. Raven records the full joint trajectory. Then, instead of you performing it 99 more times, the arm replays the motion by itself — with small variations in the trajectory each time. Slightly different paths, slightly different approach angles, slightly shifted positions.

Those small variations are the whole game. A policy trained on 100 identical episodes learns exactly one trajectory — brittle, and it breaks the moment the cup moves an inch. A policy trained on 100 varied episodes learns the task itself: the shape of the motion, the range of valid approaches, how to recover. Coverage, not repetition, is what makes a dataset good.

Toborlife AI — Unitree’s official US partner — just launched full-body teleoperation for the G1: you put on a VR headset, the robot mirrors you, and every session is captured in a training-ready format. Their own framing is telling — teleoperation as the backbone, not the product. The industry is converging on the same conclusion: the dataset is the product.

Not everyone is answering the data wall with more hardware. Mimic’s U1 exoskeleton captures human demos mapped 1:1 to their robot hand — no robot needed in the loop. Our answer goes the other way: record one clean kinematic plan, replay it with small variations, and get the coverage from software.

And here’s what learners feel immediately: it stops being tedious and starts being interesting. You demonstrate. The robot practices — dozens of variations, tirelessly — while you watch what changes and what the model picks up. You develop an intuition for what “good coverage” looks like, not from a lecture, but from watching.

Concepts before code. The dataset stops being a wall and becomes a playground.