Mass UMI data for the pretraining era of robotics
We're building the infra, ops, and tech to collect mass UMI (Universal
Manipulation Interface) data from skilled laborers in India, in order to
sell to robotics foundation model labs.
Current robots struggle to generalize to unseen situations. Simple changes
in the environment, like shifting objects around or changing the lighting,
can lead to failure. This is because robots are trained on small, narrow
data sets, usually collected via teleoperation.
UMI data collected from skilled workers and postprocessed well has comparable
quality to teleop, but it's multiple times cheaper, easier on the part of
the operator, and easier to aggressively scale because it plugs into the
existing economy.
LLMs solved the generalization problem by pretraining on the scale of all
the text humanity has ever produced (~100 million years of labor).
This gives the model a broad and deep
understanding of the world, which allows it to learn usefully from posttraining.
Pretraining is hard for robots because we don't have enough diverse data -- not
enough from teleop, not enough if you scrape every video on the internet.
But the evidence suggests it will work.
Data is the lifeblood of machine learning. Data is the bottleneck of the future.
Let's collect enough data to unlock the pretraining era of robotics.