← AI Switchboard
ROBOTS · November 11, 2026
Nov 4

Apple researchers report 91% success for a Unitree G1 humanoid trained only in a thousand simulated homes, with no human teleoperation

By the AI Switchboard · researched and written with AI from the sources listed on this page, each point checked against its source before publishing · how we work · report a mistake

The HALO system learned to sit, walk to objects, pass through doors and push boxes in generated homes, then worked in six real ones.

A paper submitted to arXiv on 4 November, “Perceptive Humanoid Motion from Simulation”, presents HALO, a system that gives a humanoid robot household movement skills learned only in simulation. Its project page lists the authors' affiliation as Apple and the robot as Unitree's G1.

The training data comes from about a thousand procedurally generated homes, not from people steering the robot. The skills include sitting, walking to a named object, passing through half-open doors, crawling under tables and pushing boxes, which the robot chains together for longer tasks.

The authors report “a 91% success rate across 360 trials in six unmapped real-world environments”, without the robot ever training in a real home. The project page itself shows a sit-task success rate of 84%.

Training robots in simulation rather than with human demonstrations is the cheapest way to scale them, if the skills carry over. Apple has said little in public about robotics; this is research, not a product.

  • Confirmed HALO reports a 91% success rate across 360 trials in six unmapped real-world environments. arXiv
  • Confirmed The project page lists the authors' affiliation as Apple, the robot as the Unitree G1, and 1,000 generated homes. HALO project page

Robots & physical AIScience & research

Catching up in the November 11, 2026 edition · front page