Xiaomi's new robot model shows data beats parameters for teaching robots to move
Robotics-1 trained on 100,000+ hours of cheap handheld-gripper footage instead of expensive teleoperation, and it beat a rival's model 75 percent to 40 percent in real world tests.
Xiaomi trained a robot model on cheap gripper footage, not costly teleoperation. It's beating rivals.
Xiaomi published Robotics-1 this week, a robot foundation model for mobile manipulation, and its main result is not a benchmark score. It is that scaling training data mattered more than scaling model size.
The data problem robotics actually has
LLMs train on effectively unlimited internet text. Robots do not have an equivalent: useful data on how to move through the physical world is expensive to collect, usually via teleoperation, a human physically piloting a robot arm for every training example. Xiaomi's team instead collected over 100,000 hours of trajectories using cheap handheld grippers, camera rigs a person can carry through a kitchen, office, or warehouse, across 1,700-plus scenarios and 2.4 million episodes, then added 7,200 hours of real robot data on top.
What that bought them
Xiaomi tested model sizes from 2 billion to 10 billion parameters and found that adding data reduced validation error more predictably than adding parameters did. Robotics-1 leads on four sim benchmarks, RoboCasa, RoboCasa365, VLABench, and RoboDojo, and in a real world unfamiliar-environment test it hit a 75 percent success rate against 40 percent for a comparable model from Physical Intelligence. Code and weights are public.
Why a build studio cares
This is the same data versus compute scaling argument that shaped the last few years of LLM development, now playing out a second time in embodied AI, with a cheaper data collection method as the actual unlock. Worth tracking for anyone evaluating robotics or physical automation work, not because the benchmark numbers are final, but because the method, cheap handheld capture instead of teleoperation, is reproducible and checkable.
Next step: see Xiaomi's technical report and the-decoder's coverage. If a physical automation or robotics angle is part of a build you're scoping, write to us at hello@gattyworks.com.