Apple discloses and open-sources HALO humanoid robot control system
Apple's research team disclosed HALO, a three-layer control system that transfers skills learned in simulation to a Unitree G1 humanoid robot, and released code for data generation, model training, task orchestration, and deployment. The system uses a Qwen-based visual-language model for task orchestration, Qwen3.5-2B with a DiT decoder for motion generation, and a reinforcement-learned tracker for joint control. During real-world deployment, a Mac mini with an M4 Pro chip runs the upper two layers onboard, while the robot's onboard hardware runs the tracker. According to the paper, HALO achieved a 91% average success rate across 360 single-task trials in six unmapped real environments and a 70% average across three composite tasks in one real environment. The team said simulated data generation still requires manually written prompts and keyframes for new physical tasks, and that limited tactile sensing and occlusion can reduce performance on contact-heavy tasks.
HALO combines synthetic training data, a deployable three-layer control architecture, and open code covering the full development pipeline, providing researchers and robotics developers with a reusable approach for transferring simulated skills to a real humanoid robot.
#Apple#HALO#Unitree G1#Qwen3.6-35B-A3B#Qwen3.5-2B#Mac mini M4 Pro#NVIDIA H100#NVIDIA B200#INFINIGEN#Isaac Sim