Xiaomi's Robotics-1 Tests VLA Scaling Laws with 100k Hours of Real Data

stepjamUK · x · 2026-08-11

The author discusses the practical relationship between data scale and model capability in robot learning, noting that previous research often disconnected pre-training scaling curves from real-world robot performance.

The review highlights Xiaomi's Robotics-1, a foundation VLA model pre-trained on over 100,000 hours of real-world manipulation data collected via handheld UMI devices. The paper is praised for testing whether scale behaves as scaling laws predict and for being honest about where those predictions break down upon contact with a real physical environment.

Original post →

More from Embodied

Embodied channel →