Apple Discloses Progress on Robot 'Brain' Research

Apple's research team submitted a paper to arXiv detailing the HALO autonomous control system for humanoid robots, validated on Unitree G1 units. The system uses a three-layer architecture combining local visual-language models (Qwen3.6-35B-A3B and Qwen3.5-2B) with reinforcement learning-based action trackers to transfer skills from simulation to real-world environments like navigating under tables or pushing boxes. The team used 250 NVIDIA H100 GPUs to synthesize training data over approximately 48 hours, though simulation data generation still involved some human input for prompts and keyframes. The system currently relies on visual perception and struggles with contact-rich tasks when objects are occluded, with tactile sensing identified as a future improvement direction.