Nov 6, 2026
Apple's HALO hit 91% on a Unitree G1 in rooms it never saw
Apple researchers trained HALO mostly on about 300 hours of sim-generated motion data. On 360 single-task trials in 6 unmapped real homes, a Unitree G1 hit 91%. Chained multi-task runs averaged about 70%.
Sir Toastimir
Video · Apple HALO project
Apple's HALO (Humanoid Autonomy, Learned in simulation, running Onboard) runs a Unitree G1 from camera input and natural-language requests, with all models onboard.
The training
The motion generator was trained mostly on about 300 hours of synthetic data: generated human-like motions placed in 1,000 procedurally generated rooms and rendered to match the robot's camera.
The results
Across 360 real-world single-task trials in six unmapped homes, HALO averaged 91% success (walk to an object, sit, step through a door, crouch, crawl, push a box). Chained multi-task sequences were harder, averaging about 70% across three tested chains.
On the robot
An orchestrator, motion generator and tracker split the work. A Mac mini M4 Pro on the backpack runs the higher-level models; the tracker runs on the G1's Jetson. Code for the pipeline is on Apple's ML GitHub; weights and the full dataset are not listed for download.
Would you let this robot crawl under your table?
Links
Get your daily slice of AI
One short email with the best new AI tools, robots and AI & tech stock news. Free. Unsubscribe anytime.
We only use your email to send the newsletter.