Apple published the HALO humanoid control system on November 7, validated with Qwen models on the Unitree G1
Apple research published the HALO autonomous control system, using Qwen3.6-35B-A3B for task orchestration and Qwen3.5-2B for motion generation, reaching a 91 percent average success rate across 360 single tasks in six unmapped real environments.
On November 7, it was reported that on November 4, Apple research published a paper on Perceptive Humanoid Motion from Simulation describing HALO, an autonomous control system for humanoid robots. The team also released data generation, model training and deployment code covering motion generation, motion tracking, task orchestration, and both simulation and real robot operation.
For deployment the team used the Unitree G1 humanoid with a custom 3D printed protective shell on the back and a Mac mini onboard, which handles real-time inference for the compute-heavy Orchestrator vision language model and the Motion Generator diffusion transformer, while the robot's own onboard chip runs the Tracker reinforcement learning policy network at high frequency. HALO is controlled by three cooperating layers: a vision language model based task orchestration module on top that defaults to Qwen3.6-35B-A3B, a motion generation layer using another Qwen model, Qwen3.5-2B, and a tracking layer trained with reinforcement learning and policy distillation that runs at 50 times per second.
The paper reports 360 single-task tests across six real environments that had not been pre-scanned or mapped, with an average success rate of 91 percent. For long-horizon compound tasks, three combinations averaged 70 percent, with the most complex sequence of crouching under an obstacle, walking to a target and pushing a box succeeding 5 times in 10 attempts for a 50 percent rate. The team used 250 NVIDIA H100 GPUs for roughly 48 hours to synthesize the main training data, and human involvement remained in the loop through noise injection and manual prompt writing for physical tasks.