Apple researchers reported that a Unitree G1 humanoid running their HALO system succeeded in 91% of 360 real-world trials, in six rooms it had never seen, after training only on simulated data (arXiv 2611.06908, submitted November 4).
The paper lists eight authors, all with Apple as their affiliation. The robot carries a Mac mini M4 Pro on its back, and the paper says the planning models run onboard. On tasks that chained two or three skills, real-world success averaged 70%, which by our arithmetic is 21 points below the single-task mean.
Business Pill · Physical AI Explained
An evergreen Business Engineer explainer on physical AI. It teaches the general idea only and does not discuss HALO or this paper.
The key insight:Apple researchers reported on November 4 that a Unitree G1 humanoid running their HALO system, trained only on simulated data, succeeded in 91% of 360 real-world trials in six rooms it had never seen, with its planning models running on a Mac mini carried by the robot.
HALO is a three-layer control stack, according to the paper and Apple’s project page. An orchestrator, an off-the-shelf vision-language model, turns a spoken request into subtasks about every two seconds. A motion generator plans five seconds of whole-body motion from the camera image about three times a second. A tracker follows that plan at 50 Hz and sets targets for 29 joints.
The hardware is a Unitree G1 with two changes, the paper says: a head-mounted camera and “a rear enclosure housing a Mac mini M4 Pro (64 GB)”. The orchestrator and the motion generator run on the Mac mini. The tracker runs on the robot’s factory Jetson Orin NX.
By default, the paper says, the orchestrator is Qwen3.6-35B-A3B running in 4-bit on the Mac mini, “to enable operation without internet connectivity”. It adds that the system “can route requests to frontier cloud models via their APIs for stronger reasoning and planning capabilities”.
Real-world success of Apple’s HALO system on a Unitree G1 in held-out rooms: six single skills (10 trials in each of six rooms) and three chained tasks (one room). Data: Hu et al. (Apple), arXiv 2611.06908, Table II. The 21-point gap is our arithmetic.
Trained in 1,000 Simulated Homes
The training data never involved a person steering the robot. The paper says the corpus is 300 hours of paired video, language and motion across 1,000 procedurally generated homes, none of which anyone visited or scanned. The project page says the corpus was generated on 250 H100 GPUs, the motion generator was trained on 64 B200s and the tracker on 4 H100s.
The paper sets this against teleoperation and wearable capture, which it says “require at least an hour of human time to produce an hour of data”. Its abstract says “The corpus scales with compute rather than human labor”. By that ratio, a 300-hour corpus collected by hand would take at least 300 hours of human time (our arithmetic).
How It Performed
The paper tested six skills: going to a named object, sitting, stepping sideways through a half-open door, crouching under a table, crawling under it and pushing a box. Each ran 10 times in each of six real rooms. Real-world success ranged from 85% for sitting to 100% for crouching, with a mean of 91%, matching the 91% mean across 100 held-out simulated homes.
Chained tasks did worse. Door then go-to succeeded in all 10 real trials, crouch then sit in 60%, and crouch, go-to and push in 50%. The chained tests used one real environment. The paper says the failures “share a single cause (the next subtask begins with its target unobserved)” and suggests adding search and turn-in-place behaviours.
As we read the paper, the product here is the software stack rather than the robot: the humanoid is a Unitree G1 with a camera and an Apple computer added.
As we read the results, single skills transferred from simulation at the same 91% mean, while chained tasks fell to 70% in the one room where they were tested.
As we read the limits section, the bottleneck has moved from collecting data to writing tasks: each new skill still needs a person to write its prompt and motion keyframes.
Hu et al. (Apple), arXiv 2611.06908
“Without ever seeing a real home during training, HALO transfers zero-shot to the physical world, achieving a 91% success rate across 360 trials in six unmapped real-world environments.”
Three Implications
FOR ROBOT MAKERS The paper says HALO ran on a Unitree G1 with two physical modifications, a head camera mount and a rear enclosure housing a Mac mini M4 Pro.
FOR AI INFRASTRUCTUREApple’s project page says the training corpus was generated on 250 H100 GPUs, the motion generator was trained on 64 B200s and the tracker on 4 H100s.
FOR BUYERS OF HUMANOID ROBOTS The paper says chained tasks failed when the next subtask began with its target out of view, and suggests adding search and turn-in-place behaviours.
The analysis says: “The transformer is a commodity now; what changes between paradigms is where intelligence accumulates and who can capture it.” It describes a Corpus Moat built by pretraining and eroding as “synthetic data closes gaps.”
As we read it, HALO is a test of that idea in robotics: the paper replaces human-collected motion data with 300 hours generated in simulation, so the scarce input becomes compute and task design. The framework does not discuss HALO or Apple.
What Is Not Established
The paper lists its own limits. Adding a new task still needs a person to write its prompt and author its motion keyframes, so “the repertoire grows only as fast as a person can specify additional tasks”. The robot sees mostly through cameras, so it “has no signal that it is still there” once an object leaves view, and contact-heavy tasks suffer. The six skills are movement skills: none involves picking up or handling objects. The chained results come from a single room. This piece makes no forecast about Apple products or timing.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Eight Apple researchers trained a humanoid’s control stack on 300 hours of simulated data from 1,000 generated homes, ran its planning on a Mac mini carried by the robot, and reported 91% success across 360 real-world trials in rooms it had never seen. Chained tasks averaged 70%.
A note on sourcing. Figures and quotations are from the Apple paper submitted to arXiv on 4 November 2026 (2611.06908) and Apple’s HALO project page. We did not reproduce the experiments. The 21-point gap and the 300-hour human-time comparison are our arithmetic. The lead reached us through an aggregator headline whose funding claim we could not source and did not use.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
