Apple unveils HALO humanoid control system with 91% real-world success rate

On November 4, 2026, Apple's research team submitted a paper titled Perceptive Humanoid Motion from Simulation to the preprint platform arXiv, disclosing HALO, an autonomous control system for humanoid robots. On November 7, Chinese tech media outlets picked up the news, adding technical details and experimental data from the paper.

HALO operates through three coordinated layers. The top layer is a task orchestration module based on a vision-language model (VLM), which defaults to Qwen3.6-35B-A3B running locally on a Mac mini housed in a custom 3D-printed protective shell mounted on the robot's back. This module shares a model serving engine with the motion generator and requires no network connection, though the system can alternatively call cloud models via API. The middle layer, the motion generation layer, uses another Qwen model, Qwen3.5-2B. The bottom layer is a motion tracker trained through reinforcement learning and policy distillation. The Mac mini, equipped with an M4 Pro chip and 64GB of unified memory, performs real-time inference for the compute-heavy Orchestrator (VLM) and Motion Generator (DiT), while the G1's onboard chip handles the high-frequency Tracker (RL policy network).

The team deployed HALO on Unitree's G1 humanoid robot, enabling it to transfer motion skills learned in simulation to unfamiliar real-world rooms, completing tasks such as crawling under tables, sitting on sofas, pushing boxes, following people, sitting on chairs, and sidestepping through half-open doors. In the real-world evaluation phase, the researchers conducted 360 individual task trials across 6 unmapped environments, achieving an average success rate of 91%. Alongside the paper, the team open-sourced its data generation, model training, and deployment code, covering motion generation, motion tracking, task orchestration, as well as simulation and real-robot operation.