Apple Unveils HALO Humanoid Robot System Powered by Mac mini
Apple researchers introduced the HALO autonomous humanoid robot control system on arXiv, leveraging a Unitree G1 hardware platform powered by an onboard M4 Pro Mac mini. The team open-sourced the complete training and deployment framework, achieving an average success rate of 91% across 360 single-task tests in six unmapped environments. This release marks a significant expansion of Apple's research into embodied artificial intelligence and robotics, demonstrating that localized consumer-grade or compact silicon like the M4 Pro can handle demanding onboard VLM workloads. Open-sourcing the data generation and deployment pipelines provides a valuable resource for the broader robotics research community. HALO uses a three-tier hierarchical architecture where an onboard Mac mini runs the high-level Vision Language Model and Motion Generator, while the robot's built-in chip executes the low-level reinforcement learning tracker at 50Hz. However, the system still relies on some manual prompt engineering and motion keyframe design for new physical tasks.
## BACKGROUND
Embodied AI combines artificial intelligence with physical hardware like robots, allowing machines to perceive, reason, and interact with the physical world. Diffusion Transformers (DiT) and Vision Language Models (VLM) serve as advanced neural architectures used to generate realistic motions and interpret complex multimodal instructions from the environment.