Apple's research team has unveiled its humanoid robot autonomous control system, HALO, on a preprint platform and deployed it on Unitree Technology's G1 humanoid robot, enabling the robot to transfer motions learned in simulation to unfamiliar real-world rooms for execution. The team has also made public the related data generation, model training, and deployment code. HALO adopts a three-tier collaborative control architecture: the top layer is a task orchestration module built on the Qwen large model, which can run locally without an internet connection and can interpret ambiguous instructions and break them down into subtasks; the middle layer uses a smaller Qwen model together with a decoder to plan full-body motions of approximately 5 seconds, updating 3 times per second; and the bottom-layer motion tracker converts motion plans into joint control commands at a frequency of 50 times per second. The research team synthesized about 300 hours of supporting training data across a thousand procedurally generated simulated home environments to complete model training, requiring only minimal human involvement. Tests show that in real environments without pre-built maps, the system achieves an average single-task success rate of 91% and an average success rate of 70% for composite tasks. Currently, the system still relies on visual perception, and its performance on contact-based tasks is easily affected; tactile perception will be added for optimization in the future.