Key Notes
- Apple’s HALO runs on a Unitree G1 and uses simulated home environments to learn physical skills before entering unfamiliar real rooms.
- The researchers report 91% single-task success across 360 trials, while a smaller evaluation of chained tasks averaged 70%.
- The work is a research system with public demonstrations, rather than an announced Apple home robot with commercial availability.
Apple researchers have introduced HALO, a system that lets a Unitree G1 humanoid respond to language instructions and move through unfamiliar indoor environments after training in simulation. The team reports 91% success across 360 physical trials covering six skills in six previously unseen environments.
The research paper, posted on November 4, describes a way to generate paired visual, language and motion data without recording every training episode through human teleoperation. HALO is a research autonomy stack running on third-party hardware; Apple has not announced a consumer robot product with a price or delivery date.
Learning Before Entering the Room
The training pipeline places generated motion into 1,000 procedurally created homes, producing approximately 300 hours of synthetic episodes. This gives the models examples of how an instruction, the robot’s camera view and an appropriate movement fit together.
The important test is transfer: the physical evaluation rooms were neither scanned in advance nor included in training. The robot had to locate relevant objects and adjust its body using onboard observations, rather than replay a route prepared for that particular room.
Three Layers Turn a Request Into Movement
On the project page, Apple shows an orchestrator choosing subtasks, a generator planning whole-body movement and a tracker translating that plan into joint commands. The generator updates roughly three times per second, while the tracker runs at 50 Hz.
This division separates the choice of a target from the mechanics of reaching it. The real-world demonstration gallery shows the G1 sitting, pushing boxes, stepping through partly open doors and lowering itself beneath furniture. Readers can watch the team’s demonstration video alongside the technical description.
What the Success Rate Measures
The 91% figure is an average over the six single-skill tests, not a measure of every activity a household robot might encounter. The paper’s separate multi-task evaluation averaged 70% across three sequences in one real environment, illustrating how longer chains can expose additional failures.
Those results are reported by the authors and have not been independently reproduced for this article. The experiments establish a bounded set of useful behaviours; they do not establish safe, unattended operation throughout an arbitrary occupied home.
Another Route Through the Robot-Data Bottleneck
RobotsBeat’s coverage of Isaac Lab 3.0 follows the simulation infrastructure behind robot learning. Our report on Tutor Intelligence examines a complementary approach built around human help and real operations.
HALO adds evidence that synthetic environments can teach perception-linked whole-body behaviour that survives the move to hardware. The next challenge is extending that transfer to more tasks and longer sequences while preserving reliable physical control.
Disclaimer: RobotsBeat is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, Robotics, technology, software, and digital innovation sectors. These relationships do not influence RobotsBeat's editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
