Apple Research Discloses HALO Autonomous Control System for Humanoid Robots Validated on Unitree G1

Apple Research disclosed the HALO autonomous control system for humanoid robots via an arXiv preprint titled 'Perceptive Humanoid Motion from Simulation'. The system was validated on a Unitree G1 robot in six unscanned real-world environments, achieving an average single-task success rate of 91%. HALO employs a three-layer architecture: a local Qwen3.6-35B-A3B visual language model for task orchestration, a frozen Qwen3.5-2B encoder paired with a DiT decoder for motion planning, and a reinforcement learning-trained tracker for joint control. Training data was synthesized from 1,000 simulated environments using 250 H100 GPUs over approximately 48 hours, though some tasks still required human prompts or motion capture assets. The system demonstrated tasks like crawling under tables and pushing boxes but currently lacks tactile perception.