Researchers at Apple have published details of an autonomous control system for humanoid robots called HALO, according to a paper submitted to the arXiv preprint repository on 4 November 2026.
The team tested the system on a Unitree Robotics G1 humanoid robot, enabling it to execute movements learned in simulation inside unfamiliar real rooms. HALO uses a three-tier setup: a top-level vision-language model for task orchestration running locally on a Mac mini, a middle layer generating five-second whole-body motion plans three times per second, and a low-level motion tracker running at 50 hertz to convert plans into joint commands.
In tests across six unscanned real environments, the system completed 360 individual tasks with an average success rate of 91 per cent. In a single room testing three combined task sequences, the average success rate reached 70 per cent, including a 50 per cent success rate for crawling under an obstacle, walking to a target and pushing a box.
The team generated about 300 hours of synthetic training data across 1,000 simulated home environments using 250 Nvidia H100 graphics processing units over 48 hours. The researchers noted that the system currently relies on visual sensing and struggles when objects leave camera view, with tactile sensing planned for future development.
