Apple HALO: Unitree G1 humanoid trained only in simulated homes crawls under tables in real ones
Apple machine learning researchers have published HALO, which teaches a Unitree G1 humanoid household skills entirely in simulation, using about a thousand procedurally generated virtual homes. The dataset grows with computing power, not human teleoperation time.
Source: XRoboHub on X
In videos Apple labels fully autonomous, the robot sits on chairs, walks to named objects, sidesteps through half-open doors, crawls under tables and pushes boxes, chaining skills from plain-language requests. It carries a head camera and a backpack Mac mini and runs untethered.
Why it matters
The authors report 91% success across 360 single-skill trials in six real homes it never saw; multi-step tasks scored lower. The Nov. 4 preprint hasn't been independently checked. Humanoid training often relies on human demonstrations that cost an hour of human time per hour of data, so a sim-only recipe, with code released, could cut that cost if it holds up.
More on Unitree and humanoid robots.
About the footage: Autonomous, per Apple's project page: official HALO gallery clip of the G1 crawling under a table, trimmed to 3.5 s and muted
The source post
Post by Apple Machine Learning Research on X.
View the post on X
Sources
New to robots? Read our guide: How humanoid robots work.
Summarized by Servospace from the sources above, drafted with AI help and checked against those sources before publishing. Claims are the companies' or outlets' unless we say otherwise. How we work.