Apple's HALO Humanoid Control System Hits 91% Success on Unitree G1

On November 4, Apple's research team submitted the paper “Perceptive Humanoid Motion from Simulation” to the arXiv preprint platform, unveiling HALO, an autonomous control system for humanoid robots.

The team deployed HALO on Unitree Robotics' G1 humanoid robot to verify whether motions learned in simulation could be carried into unfamiliar real-world rooms. The report notes that the validation involved invoking the Qwen model.

HALO reached an average success rate of 91% across cross-scene tasks, according to the report. The team says the results indicate that simulation-trained control policies can transfer to real environments with high reliability, a step toward deploying humanoid robots in diverse settings.

The system is designed to help humanoid robots generalize beyond their training environments, addressing a long-standing challenge in robotics. The paper details a perceptive motion approach that combines simulation training with real-world sensing to support autonomous operation.

Apple's disclosure adds to growing industry efforts to test humanoid robots in realistic conditions before commercialization.