Physix Frontier · News Briefing Card (IT Home · Nov 7, 2026)

Apple unveils HALO humanoid system, 91% success in unmapped spaces

KEY FACTS

Apple's research team submitted a paper to arXiv on November 4, publicly detailing the HALO humanoid robot system.
HALO uses a three-tier control architecture: a top-level VLM orchestrates tasks, a mid-level layer generates motions, and a low-level tracker runs at 50Hz.
The team completed 360 single-task tests across 6 unmapped real-world environments, achieving an average success rate of 91%.
For long-sequence composite tasks, three combinations were each tested 10 times, averaging a 70% success rate, with the most complex combination reaching only 50%.
The research used 250 H100 GPUs to synthesize training data, and trained the motion generator on 64 B200 GPUs.

KEY DATA

91%Average single-task success rate

6Number of unmapped test environments

70%Average composite-task success rate

50%Success rate of most complex composite task

PHYSIX OBSERVATION

Apple running hierarchical control on a Unitree G1 plus a Mac mini shows the humanoid robot race is shifting from hardware to data and model pipelines. The 91% single-task success rate is impressive, but composite tasks falling to 70% or even 50% exposes shortcomings in long-sequence planning and the lack of tactile sensing. Open-sourcing the code and the simulation data generation path could lower the barrier for the industry to reproduce the work.

Source: IT Home report