Apple unveils halo humanoid system, achieving 91% success rate in unmapped environments

Apple research teams recently published a paper on arXiv detailing HALO, an autonomous control system for humanoid robots. The researchers also released the associated code for data generation, model training, and deployment, covering aspects like motion generation and task orchestration. For real-world testing, the system utilized a Unitree G1 robot equipped with a custom 3D-printed shell and a Mac mini running an M4 Pro chip. HALO operates on three control layers: a Vision-Language Model (VLM) for task planning (defaulting to Qwen3.6-35B-A3B), a Qwen3.5-2B model for motion generation, and a low-level tracker running at 50Hz via reinforcement learning. Testing across 6 unmapped environments showed an average success rate of 91% for single-task tests. For complex, multi-step tasks, the average success rate across three combinations was 70%, with the most difficult sequence achieving 50% success. The training process required 250 NVIDIA H100 GPUs for data synthesis and involved training the motion generator with 64 B200 GPUs and the tracker with 4 H100s. The team noted that future improvements will focus on integrating tactile sensing, as the current system struggles with contact-based tasks when objects leave the camera's view.