Essays · July 30, 2026

Notes on Xiaomi-Robotics-1

These are my notes on Xiaomi-Robotics-1, a vision-language-action model trained on more than 100,000 hours of real-world manipulation data. The result that matters is not the benchmark scores but that manipulation now scales the way language models do: performance improves predictably with more data and a bigger model, the gains hold up on real hardware, the binding constraint is data rather than model size, and the data has become cheap to collect. Four points, one chart each, and the tiers are orders of magnitude, not benchmarks.

1 · Scaling laws come to robotics

Both levers move the same way: more data raises success, and a bigger model raises it again, both along a smooth curve, not in steps. That is how large language models scaled, and I did not expect it to hold this cleanly for something as physical as moving objects around a room.

More data and a bigger model both pay off predictably Out-of-box success (%), higher is better 26 53 75 more data → 61 75 79 bigger model → Predictable scaling the LLM curve, now physical More training scale, further right is more Values (26 / 53 / 75 and 61 / 75 / 79) are the paper's reported figures. Positioning illustrative.
Both levers move success up the same way: add data, add parameters, get more skill. That smooth, repeatable rise is what a real scaling law looks like, not a one-off tuning win.

2 · It transfers to the real world

The usual failure mode is a model that scores well on a benchmark and then fails in an actual home. That does not happen here: a model that is stronger in pre-training is a better robot on real hardware, in homes it never saw during training, so the benchmark number and the real-world number move together.

A better model in the lab is a better robot in a stranger's home Success on real robots in unseen homes, higher is better weaker model stronger model The gains transfer stronger pre-training, better in unseen homes Pre-trained model strength, further right is stronger Illustrative of the reported relationship; the curve shape is not a benchmark.
The pre-training work shows up where it counts, on real hardware in places the robot has never been, not only on a benchmark.

3 · The bottleneck is data, not model size

Past a few billion parameters, a bigger model barely helps and the curve flattens, while more data, and more varied data, keeps paying. That changed how I think about where the effort goes: if I were allocating against this, I would be buying data, not a bigger model.

More parameters stall; more data still pays Success, higher is better more data → Data still climbing, the binding constraint 61 75 79 bigger model → Model size capacity already enough More of each input, further right is more Model points 61 / 75 / 79 are reported (2B / 5B / 10B). Data curve illustrative.
Same outcome, two very different slopes: capacity has saturated, data has not, so the money is better spent collecting data than growing the model.

4 · And the data is now cheap

Collecting manipulation data no longer needs a robot: a UMI handheld gripper lets anyone record it, and a vision-language model labels each clip automatically. That is what makes the rest practical, turning "collect more data" into a plan, not an aspiration.

Take the robot and the human out, and data gets cheap Data throughput and scalability, higher is better the shift Old pipeline robot + teleoperation + manual labels New pipeline UMI gripper + VLM auto-labelling cheap · parallel · 100k+ hrs, no robot ~ cheap ~ $$ ~ $$$ Cost per hour of data, further right is pricier Illustrative positioning; the 100k+ hrs figure is reported.
Take the robot and the human labeller out of the loop and data collection stops being a capex problem. That cheap, parallel data is the input feeding every chart above.

Put together, it looks like a feedback loop. Manipulation scales predictably, the gains survive the real world, the constraint is data rather than model size, and the data is cheap to collect at scale. If that holds, the advantage shifts from owning the biggest model to owning the largest and most varied dataset. That is the version of a general-purpose robot policy that reads like a plan rather than a promise. I would want it replicated outside Xiaomi before relying on it, but the direction is clear.

Figures from Xiaomi-Robotics-1 (arXiv 2607.15330). Charts 2 to 4 use illustrative curve shapes; the reported numbers are the panel-1 values. The framing and any errors are mine.