Apple's HALO Robot System Hits 91% in 360 Trials on Unitree G1

Apple researchers reported a 91% success rate across 360 single-task trials for a humanoid-robot control system called HALO, tested in six real rooms the machine had never seen. The paper, "Perceptive Humanoid Motion from Simulation," landed on arXiv on November 4, 2026. The test platform was a commercially available Unitree Robotics G1.

The most striking detail is the compute. AppleInsider reported that the robot carried an M4 Pro Mac mini with 64GB of unified memory on board. All inference ran locally, not in a data center. Reporting on the disclosed system identified the task-planning model as Alibaba's Qwen3.6-35B-A3B.

What did Apple actually disclose?

The verified record is short. Six real environments, none pre-scanned or mapped. Three hundred sixty single-task trials at 91% average success. Tasks included obstacle navigation, box pushing, person following, passing through narrow openings and sitting on furniture. A combined test chained three task types, 10 runs each in one environment, and averaged 70% across 30 trials.

Training used roughly 300 hours of simulated human-like motion across about 1,000 virtual rooms. The robot's height, weight, joint count and price were not disclosed.

That is the whole confirmed list. It describes a research result, not a product. Nothing gives a launch date, a price or a robot of Apple's own design.

What is unconfirmed in the coverage?

Three claims in circulation lack evidence:

• A decade-long secret R&D program. No source documents a program of that length or secrecy.
• HALO as a consumer product. No sales plan or product announcement exists.
• A strategic Unitree partnership. Only use of the G1 platform is confirmed.

Chinese outlet 36Kr ran the story under a headline claiming Apple's decade-long secret humanoid "Finally Adopts Unitree Technology." The paper says no such thing, and neither does any other reviewed source. Using a robot for an experiment is not building a technology alliance.

How does the training method work?

HALO learns perceptive humanoid motion from simulation. The team generated training data inside virtual spaces instead of filming humans or teleoperating a robot for months. The volume: roughly 300 hours of synthetic motion across about 1,000 virtual rooms.

Real-world robot data is slow and expensive. Each hour of teleoperation costs an hour of paid time plus machine wear. Simulation removes that ceiling, regenerating room layouts in seconds. The catch is the sim-to-real gap: lighting, floor friction and doorway positions never match a model's expectations exactly. HALO's result suggests the gap can be narrow enough for unfamiliar homes and offices.

What do the numbers mean in practice?

Simple arithmetic puts the headline in scale. At 91% over 360 trials, roughly 328 succeeded and 32 failed. That is a low failure rate for a humanoid in unfamiliar space, but not one that permits unsupervised kitchen work. One failure in ten means a dropped box, a missed doorway or a stalled sit.

The 70% combined-task figure tells the more realistic story. When the robot chained three task types, success dropped 21 percentage points. Errors compound: at 90% per step, three independent steps succeed about 73% of the time, close to the reported 70%. A twelve-step chore chain at 91% per step would succeed under a third of the time. The disclosure does not describe recovery behavior, so that gap remains open.

The reporting gives averages only, not per-task breakdowns. A strong mean can mask a weak task such as narrow openings, offset by easy ones such as person following.

Why a Mac mini and an Alibaba model?

Unified memory is the key. On Apple silicon, CPU and GPU share one pool, so a 64GB pool holds a mid-sized language model plus perception and control software simultaneously. Local compute also removes network round trips. A humanoid waiting on a cloud model stumbles; onboard compute keeps the loop tight and keeps camera data private. The Mac mini is a desk-sized box, though, and reads as a prototype brain, not a production preview.

Apple's researchers did not build the planner from scratch. The "35B" label refers to total parameters, and "A3B" conventionally signals about 3 billion active per token, a mixture-of-experts pattern that keeps inference cheap. Splitting planning from motion control lets researchers swap planners without retraining the body controller.

How did markets react?

Chinese robotics stocks surged after the research, according to Yahoo Finance UK. The confirmed link is a test on a Unitree G1 and nothing more. Traders who treated the paper as a supply deal priced an unconfirmed claim. The disciplined takeaway: technical validation, not a revenue event.

What does this change for Apple's silicon strategy? It demonstrates that the company's chips can run a demanding, real-time, multi-model workload on a moving machine — a stress test few consumer scenarios match, and a vivid argument for unified memory in local AI.

What to watch next

Three things will clarify the picture: whether Apple or Unitree comment on their relationship; whether the team releases code, data or per-task numbers; and whether follow-up work moves to the longer task chains that real chores require. Until then, the summary stays simple: simulated training, a Mac mini planner, 91% of 360 single tasks on a Unitree G1. A real result — not proof of a secret program, a consumer robot or a partnership.