Apple Built a Robot Brain and Tested It on a Unitree
HALO runs open Qwen models on a Mac mini strapped to a G1, and the results are unusually specific.
Apple researchers posted a paper titled Perceptive Humanoid Motion from Simulation to arXiv on 4 November 2026, describing HALO, a control system tested on a Unitree G1 fitted with a 3D-printed housing holding a Mac mini with an M4 Pro chip and 64GB of memory. The top layer, a vision-language planner, defaults to Qwen3.6-35B-A3B, while the motion generator uses Qwen3.5-2B and a learned tracker runs at 50 times per second. Across six unmapped rooms and 360 single-task trials, average success reached 91%; three-task chains in one room averaged 70%, and the hardest chain succeeded five times in ten.
