Apple Research Team Discloses HALO Autonomous Control System for Humanoid Robots

On November 4, Apple's research team submitted a paper to arXiv detailing the HALO autonomous control system for humanoid robots. Validated on Unitree G1 units in six unscanned real-world environments, the system achieved an average success rate of 91% in single-task tests and 70% in combination tasks. HALO employs a three-layer architecture: a visual language model (VLM) for task planning, a frozen encoder with a DiT decoder for whole-body action generation, and a reinforcement learning-trained action tracker for joint control. The system was trained using synthetic data generated from 1,000 simulated environments, involving approximately 300 hours of paired instruction-visual-action data. The research also notes limitations in tactile perception and plans to address them in future iterations.