Apple-affiliated researchers publish HALO preprint claiming 91% zero-shot success transferring simulated humanoid skills to real homes

Apple-affiliated researchers have released a preprint, code and videos for a humanoid robot system they call HALO, saying it learned entirely in simulation and then transferred directly to real homes with no real-home training data. In the paper’s headline result, the authors report “achieving a 91% success rate across 360 trials in six unmapped real-world environments.”

The work was posted to arXiv on Nov. 4, 2026, as “Perceptive Humanoid Motion from Simulation,” or arXiv:2611.06908v1. The listed authors are Siyuan Hu, Abulikemu Abuduweili, Yue Meng, Shipeng Liu, Quentin Leboutet, Chen Chen, Matthias Müller and Vladlen Koltun. The project page is hosted by Apple Machine Learning Research, identifying the work as Apple-affiliated.

Why this matters is straightforward: humanoid robotics has long faced a data bottleneck. Training a robot to turn camera input and natural-language commands into full-body movement usually requires expensive paired data, often gathered through human teleoperation — where a person remotely controls the robot — or motion-capture systems that record human movement. HALO’s central claim is that much of that labor can be replaced with synthetic data generated in simulation.

In the arXiv abstract, the researchers say HALO builds vision-language-motion data across 1,000 procedurally generated homes rather than relying on those traditional collection methods. As the abstract puts it, “The corpus scales with compute rather than human labor.” That is a notable claim in a field increasingly interested in simulation and large motion datasets, especially because this project combines vision, language and full-body humanoid control and reports zero-shot real-world transfer, meaning the system is tested in real settings without task-specific real-world training.

According to the abstract, HALO learned skills including sitting on chairs, navigating to named objects, stepping sideways through half-open doors, crouching and crawling beneath tables, and pushing boxes. The authors also say the system can combine those abilities into longer, multi-step tasks rather than handling each behavior in isolation.

The release is unusually open by robotics standards. The Apple project page says code and videos are publicly available, and the public repository, apple-aiml-research/ml-halo, includes code for data generation, model training and deployment. The project materials show the system running on the Unitree G1, a commercial humanoid robot platform.

The project page also gives a sense of the scale behind the synthetic-data pipeline. It describes the training world as “1,000 homes,” and says, as an author-reported detail, “The corpus is generated on 250 H100.” H100 refers to Nvidia’s high-end AI chips, a reminder that replacing human data collection with simulation does not make the process cheap so much as shift the cost toward computing infrastructure.

The paper’s most attention-grabbing figure, however, should be read with caution. This is a preprint on arXiv, not a peer-reviewed paper, and the reported performance numbers are author claims from the paper and project materials. As of Nov. 5, 2026, there was no independent public validation of the 91% result.

That caveat matters because sim-to-real transfer — moving from simulation to the physical world — has been a stubborn challenge in robotics. Success in simulation often fails to carry over cleanly to real environments, especially for humanoids operating in cluttered spaces with stairs, doors, furniture and narrow passages.

Still, if HALO’s results hold up, the release points to a practical route around one of humanoid robotics’ biggest constraints. Instead of collecting vast amounts of human-demonstration data by hand, researchers may be able to generate training data at scale in simulation and then deploy those policies on real robots with far less real-world data collection.