Apple Unveils Humanoid Robot System HALO, Powered by Two Qwen Models
On November 4, an Apple research team submitted a paper to arXiv, unveiling HALO, an autonomous control system for humanoid robots. Verified on a Unitree G1, HALO executes actions learned in simulated environments within unfamiliar real-world rooms. HALO features a three-tier architecture: the top tier uses Qwen3.6-35B-A3B for task orchestration, the middle tier employs Qwen3.5-2B to generate actions, and the bottom tier is a reinforcement learning-trained action tracker. The team has also…
Apple Unveils Humanoid Robot System HALO, Powered by Two Qwen Models
On November 4, Apple's research team submitted a paper to the preprint platform arXiv, revealing its humanoid robot autonomous control system, HALO. The system was validated on a Unitree G1 robot, successfully executing actions learned in a simulated environment within an unfamiliar real-world room.
HALO is structured in three layers: the top layer uses Qwen3.6-35B-A3B to orchestrate tasks, the middle layer employs Qwen3.5-2B to generate actions, and the bottom layer is an action tracker trained through reinforcement learning. The research team has also released the code for data generation, model training, and deployment.
Sina Tech