Shattered | 2026-11-13

Apple researchers have put a humanoid robot through six real rooms it had never seen, and the machine finished 91% of its single tasks. The system is called HALO, the paper is titled “Perceptive Humanoid Motion from Simulation,” and it landed on arXiv on November 4, 2026. AppleInsider reported that the robot carried an M4 Pro Mac mini on board, which means the Apple HALO humanoid robot did its thinking locally rather than in a data center.

The hardware under the experiment is a Unitree Robotics G1, a commercially available humanoid. That detail drove most of the headlines. One outlet, 36Kr, ran the story under a title claiming that Apple’s decade-long secret R&D humanoid robot “Finally Adopts Unitree Technology.” The paper itself does not say that, and neither does any other source we reviewed. What it shows is narrower and, in some ways, more interesting.

This analysis separates what Apple has actually disclosed from what the coverage has implied. It covers the training method, the test numbers, the compute stack, the market reaction and the competitive picture. Readers who follow Apple’s local AI push may also want our earlier look at the Harbor app that would pool memory across Macs, since the HALO setup leans on the same idea of serious on-device inference.

The confirmed record is short and specific. Apple researchers described a humanoid-robot control system named HALO in a paper submitted to arXiv on November 4, 2026. The system was tested on a Unitree Robotics G1 humanoid. Tests ran across six real environments that nobody had pre-scanned or mapped for the robot.

Across 360 single-task trials, HALO reached an average success rate of 91%. The reported tasks included navigating around obstacles, pushing boxes, following a person, passing through narrow openings and sitting on furniture. A separate evaluation chained tasks together. That combined test averaged 70% across 30 trials, with three task types run 10 times each in a single real environment.

The robot carried a Mac mini with an M4 Pro chip and 64GB of unified memory. Reporting on the disclosed system identified the local task-planning model as Qwen3.6-35B-A3B, developed by Alibaba. Training reportedly used about 300 hours of simulated human-like motion spread across roughly 1,000 virtual rooms.

That is the whole verified list. It describes a research result, not a product. Nothing in the disclosure gives a launch date, a price, a retail plan or a robot of Apple’s own design.

The headline version of this story contains three claims the evidence does not support. The first is that Apple has run a decade-long secret humanoid-robot program. No source in our verified set confirms a program of that length or that secrecy. A single research paper cannot establish ten years of hidden work, and Apple has not described one.

The second claim is that HALO is a consumer product, or that Apple plans to sell a humanoid. Nothing disclosed points that way. Research groups at large companies publish papers constantly, and most never become products.

The third claim is the one in 36Kr’s title: that Apple adopted Unitree technology as a strategic partner. The available reporting confirms use of a Unitree G1 as the test platform. It does not confirm a partnership, a supply agreement or any formal adoption. Buying or borrowing a robot to run an experiment is a different act from building a technology alliance.

The title of the paper points to the core idea. HALO learns perceptive humanoid motion from simulation. Instead of filming people walking through real houses or teleoperating a robot for months, the team generated training data inside virtual spaces. Reporting puts the volume at roughly 300 hours of simulated human-like motion across about 1,000 virtual rooms.

Those numbers matter for a practical reason. Real-world robot data is slow and expensive to collect. A human operator wears a motion suit or holds a controller, and each hour of useful footage costs an hour of paid time plus the wear on a machine. Simulation removes that ceiling. A room can be regenerated in a new layout in seconds, and a thousand rooms give the model a thousand different furniture arrangements to fail in before it ever touches a real floor.

The catch is the gap between a virtual room and a real one. Lighting differs, floors have friction the simulator approximates, and doorways are never quite where a model expects. Researchers call this the sim-to-real gap, and closing it has been the central struggle of robot learning for years. HALO’s headline result is evidence that the gap can be narrow enough to work in homes and offices the system never saw.

A 91% average over 360 single-task trials is a strong headline, but it deserves some arithmetic. The figures below come from the reported numbers, plus simple multiplication. The derived rows are our calculation and not figures Apple published.

Roughly 32 failures out of 360 is a low rate for a humanoid moving through unfamiliar space. It is not a rate that lets a robot work unsupervised in a kitchen. One failure in ten means a dropped box, a missed doorway or a stalled sit. For a lab demonstration, that is impressive. For a household appliance, it is far short.

Readers should also note what an average hides. The reporting gives a mean across tasks and rooms, not a per-task or per-room breakdown. A strong average can mask a weak task, such as squeezing through a narrow opening, offset by easy ones such as following a person. Until the full paper tables are examined task by task, the 91% should be read as a headline figure and not a guarantee for any single behavior.

The 70% figure tells a more realistic story than the 91%. When the robot had to chain three task types in one real environment, average success dropped by 21 percentage points. Thirty trials is a small sample, so the number carries wide uncertainty, but the direction is what you would expect.

Errors compound. If each step succeeds nine times in ten and the steps are independent, three steps in a row succeed about 73% of the time. That back-of-envelope figure sits close to the reported 70%, which suggests the drop comes mostly from stacked small failures and not from a separate weakness in planning. This is our reading of the arithmetic, not a finding from the paper.

Real chores run much longer than three steps. Clearing a table, carrying items to a counter and loading a dishwasher might involve a dozen. At 91% per step, a twelve-step chain would succeed less than a third of the time. That is why the field cares so much about recovery behavior, meaning the ability to notice a failure and retry. The disclosure does not describe how HALO handles recovery, so we cannot say how far it has come on that front.

The most surprising detail is the compute. The robot carried a Mac mini with an M4 Pro chip and 64GB of unified memory. AppleInsider headlined the setup as a Mac mini powering a staggeringly capable robot, and the choice of an off-the-shelf desktop-class machine says a lot about Apple’s priorities.

Unified memory is the key. On Apple silicon, the CPU and GPU share one pool, so a large model does not need to be copied between separate memory banks. A 64GB pool is enough to hold a mid-sized language model plus the perception and control software at the same time. That makes the Mac mini a convenient inference box, and it explains why Apple researchers would reach for one before designing custom robot hardware.

Running locally also removes network round trips. A humanoid that waits on a cloud model to decide whether to step left will stumble, and a robot in a house with weak Wi-Fi is worse off still. Onboard compute keeps the loop tight and keeps camera data inside the machine. If you are weighing similar hardware for your own local AI projects, our Mac mini M6 vs iMac vs MacBook Pro comparison lays out the price tiers.

The Mac mini also has an obvious limit. It is a desk-sized box with its own power supply and cooling needs, strapped to a robot for a research test. Nobody should read the setup as a preview of Apple’s production design. It reads as a convenient prototype brain.

Reporting on the disclosed system identified the local task-planning model as Qwen3.6-35B-A3B, developed by Alibaba. That is a notable detail for a company known for guarding its stack. Apple’s researchers did not build the planner from scratch. They used a model that anyone can study, and ran it on Apple hardware.

The name follows a naming convention common to this model family. “35B” refers to total parameters, and “A3B” conventionally signals that only about 3 billion are active for any single token, which is how mixture-of-experts designs keep inference cheap. We read the name that way but the paper reporting does not spell out the architecture, so treat it as an inference from the label.

The planner’s job is to turn a spoken or typed goal into a sequence the motion system can execute. HALO’s motion layer handles balance, footsteps and contact. The language model decides what to do next. Splitting the work this way lets researchers swap planners without retraining the body controller, which is probably why a third-party open model was practical here.

For readers tracking open models, this is a data point worth filing. Our roundup of open source AI models and the three ways to run them covers the wider field. The HALO stack shows that open-weight models are now good enough for a major lab to use as a component in a robotics paper.

Choosing a Unitree Robotics G1 is a pragmatic move. Building a humanoid is expensive and slow. A commercial platform lets a software team focus on the control and perception problem and skip the actuator design. Many academic labs use commercial humanoids for the same reason, and publishing results on a common robot makes the work easier for others to reproduce.

The reporting does not disclose the G1’s height, weight, joint count or price in Apple’s setup, so we will not guess at them. What matters is the claim the choice supports: HALO’s method is not tied to a custom Apple body. If the same approach works on a purchasable robot, then other teams can test it too.

It also explains the gap between the facts and the framing. Using a robot is not adopting a company’s technology. The distinction matters for investors and readers, because an adoption claim implies commercial ties and a research-platform claim does not. Until Apple or Unitree says otherwise, the accurate description is that Apple researchers validated their software on a Unitree G1.

Markets responded quickly. A Yahoo Finance UK item was headlined around a surge in Chinese robotics stocks following Apple’s robotics research. We have not verified the size of any individual move, so we do not quote percentages here.

The pattern is familiar. A big-name company touches a supplier or a category, and traders rush to price in a partnership that may not exist. As the table above shows, the confirmed link is a test on a Unitree G1 and nothing more. Traders who treated the paper as a supply deal were pricing an unconfirmed claim.

Capital is already flowing into the category for stronger reasons. Our report on Agile Robots raising $500M at an $8.5B valuation shows investors backing robot makers directly, and the broader hardware coverage on this site tracks how that money reaches chips and components. A paper from Apple adds a signal that a very large platform company is paying attention, even if it has not announced a plan.

The takeaway for investors: treat the paper as technical validation, not a revenue event. Apple has not announced a robot, and the Unitree link is a test platform. A rally built on partnership expectations can reverse fast when those expectations go unmet.

HALO sits inside a wider argument about how humanoid robots should learn and where they should think. The table compares three broad approaches. It is our qualitative framing of the design space, not a ranking of specific companies, and it makes no claims about any rival’s results.

HALO’s reported design bets on the first row. Its appeal is cost. A thousand virtual rooms cost compute time, not hardware wear. Its risk is the gap between virtual and real, which the 91% result suggests is manageable for the tested tasks.

Other players lean on different rows. Our coverage of the Optimus Academy claim of 10,000 bots reflects the fleet-and-training route, where scale of deployed machines is the lever. Apple’s paper offers a counterpoint: a lab can reach respectable real-world performance without a large fleet if simulation is good enough.

Robotics researchers have fought the sim-to-real problem for years. Early simulators were too crude to transfer anything useful, and controllers that looked great in software fell over on contact with a real floor. Progress came from better physics engines, randomization of simulated conditions and faster GPUs. Each step made synthetic practice a little more trustworthy.

Humanoids made it harder. A wheeled robot can fail gently. A two-legged robot falls, and a fall can break a very expensive machine. Perception adds another layer, because a humanoid has to see furniture, judge distances and plan footsteps at the same time. HALO’s contribution, as described, is a pipeline that ties perception, language and motion together using generated data.

Language models changed the picture again. Once a model could read a goal in plain words and break it into steps, the planning layer stopped being a bottleneck. That shift made it plausible to pair a language planner like Qwen3.6-35B-A3B with a motion controller trained in simulation. The pieces existed separately. HALO shows them working together on a commercial humanoid.

Research papers rarely map one-to-one to launches. Because the fact sheet confirms no robotics product and no long-running program, the safest framing places HALO among the many research efforts that inform a company’s thinking without committing it to a market.

Even without a product, the experiment has strategic value for Apple. It demonstrates that its chips can run a demanding, real-time, multi-model workload on a moving machine. Running a language model, perception and motion control together is a stress test that few consumer scenarios match.

It also extends the case for unified memory in AI work. A 64GB pool that serves both the planner and the perception stack is the sort of design advantage Apple likes to show off. Anyone benchmarking local inference on a Mac mini will recognize the pitch, and a robot is a vivid way to make it.

None of this proves Apple will build a robot. It does suggest the company’s researchers are probing embodied AI seriously enough to publish. For hardware watchers, the lesson is that the chip, the memory and the on-device model matter more than the shape of the body that carries them.

The following are our own forecasts, based on the confirmed record. None come from Apple or any named source.

Tactile sensing is one area where we expect pressure. HALO’s reported tasks rely on vision and motion, while touch remains a gap. Open efforts such as the LingBot-Touch tactile model trained on 5,000 hours point to where complementary research is heading.

Three things will clarify the picture. First, whether Apple or Unitree comment on the relationship, which would settle the adoption question. Second, whether the research team releases code, data or a longer set of results, since per-task numbers would show where the 91% is strong. Third, whether follow-up papers move from single and triple task chains to the longer sequences that real chores require.

Until then, the disciplined summary is simple. Apple researchers built a system that learned from simulated rooms, ran its planner on a Mac mini, and completed 91% of 360 single tasks on a Unitree G1 in six unseen environments. That is a real technical result. It is not proof of a secret decade-long program, a consumer robot or a partnership.

What is Apple HALO? HALO is a humanoid-robot control system described by Apple researchers in the paper “Perceptive Humanoid Motion from Simulation,” submitted to arXiv on November 4, 2026. It learns motion from simulated environments and was tested on a real robot.

Which robot did Apple use? The tests ran on a Unitree Robotics G1 humanoid. The reporting does not give the robot’s height, weight, joint count or price in the experiment, so those details remain unconfirmed.

Did Apple partner with Unitree? No partnership has been confirmed. Reporting confirms only that Apple researchers used a Unitree G1 as the test platform. 36Kr’s headline described the story as Apple adopting Unitree technology, but the evidence supports platform use alone.

How well did HALO perform? It averaged a 91% success rate across 360 single-task trials in six real environments that had not been pre-scanned or mapped. A combined-task test averaged 70% across 30 trials in one real environment.

What hardware ran the system? The robot carried a Mac mini with an M4 Pro chip and 64GB of unified memory, according to AppleInsider and other reports. The task-planning model was reportedly Alibaba’s Qwen3.6-35B-A3B.

Is Apple planning to sell a humanoid robot? There is no confirmation of that. The disclosure is a research paper, and no product, price or launch date has been announced.

Was this a decade-long secret program? That claim is unconfirmed. The paper establishes a recent research result and says nothing verified about a program running for ten years.

How was HALO trained? Reporting says the team used roughly 300 hours of simulated human-like motion across about 1,000 virtual rooms, then tested the result in real spaces the system had never seen. Asia Daily and BitcoinVersus.Tech also covered the 91% result across 360 unseen-room trials.

Read more