Apple’s humanoid runs Qwen on a Mac mini and clears 91% of real-room trials

Apple researchers have built a system called HALO that lets a Unitree G1 humanoid robot sit, crouch, crawl under tables and walk to named objects using a Mac mini carried on its back. The work, described in an arXiv paper dated November 4, reports a 91% success rate across 360 trials in six real environments the robot had not mapped.

A Mac mini strapped to a humanoid

The paper, titled “Perceptive Humanoid Motion from Simulation,” lists eight authors from Apple, including Vladlen Koltun. The hardware is a Unitree G1 with a custom head mount holding an OAK-1W camera and a rear enclosure for a Mac mini M4 Pro with 64 GB of memory. A Jetson Orin NX handles tracking, and a downward-facing Intel RealSense D435i supplies depth.

The software is split into three layers. An orchestrator, Qwen3.6-35B-A3B in 4-bit MLX format, runs onboard on the Mac mini and is invoked every two seconds to decide what to do next. A motion generator, built from a frozen Qwen3.5-2B encoder and a diffusion transformer decoder of roughly 49.3 million parameters, produces motion plans at about 3 Hz. A tracker then turns those plans into joint commands at 50 Hz. The authors say running everything onboard lets the robot work without an internet connection.

Training happened entirely in simulation. The paper describes about 300 hours of episodes across a thousand procedurally generated homes. The appendix gives 162,698 episodes totaling 296.3 hours, and the “homes” are generated single rooms such as living rooms, offices and bedrooms. Six skill datasets cover going to an object, sitting on chairs, stepping through half-open doors, crouching, crawling beneath tables and pushing boxes.

Single skills work, chains of skills struggle

The 91% headline is a mean over single tasks, with 10 real trials per task in each of six environments. According to the paper’s Table II, real-world success was 92% for going to an object, 85% for sitting, 93% for doors, 100% for crouching, 90% for crawling and 88% for pushing.

Chained tasks fare worse. In real trials, door followed by go-to reached 100%, but crouch then sit scored 60% and a three-step crouch, go-to and push sequence scored 50%, for a multi-task mean of 70%. The authors write that chained success falls below the product of the single-task rates, and note that some multi-step failures begin with a target out of view.

What the numbers say about the design

The paper’s own figures show where the gains come from. The simulation benchmark, 6,000 single-task trials over 100 held-out environments, also averaged 91%, and multi-task simulation averaged 69%. Real and simulated results are nearly identical, which suggests the randomized visual conditions used in training, such as blur, low light and color tints, carried over to physical rooms. That is the authors’ framing of the result; the paper does not report a test of it on other robots.

The ablations are more telling. Removing depth input drops sitting success from 84% to 74%. Replacing the flow-matching diffusion transformer with a plain regression decoder drops success to 19%, according to Table III. That is a large gap for a component swap, and it supports the case for generative motion models over simple regression for humanoid control.

On cost, the robot itself is no longer exotic. Unitree announced the production G1 at about $16,000 in 2024, according to The Robot Report, though current prices vary by seller. The Mac mini M4 Pro is an off-the-shelf desktop, so the compute side of the stack is also purchasable hardware. The paper does not give a total bill of materials, and the orchestrator is open Qwen weights rather than a proprietary Apple model.

The Qwen choice also fits a pattern of Qwen models turning up in research setups; our recent story on how GPT-6 Astra edited a small Qwen model is a different use of the same family.

What the paper does not settle

The results are the authors’ own, from a preprint that, as far as we could see, has no head-to-head success-rate comparison with prior systems; its comparison with earlier work is a qualitative table of capabilities and trial counts. The trials took place in six environments, which is a small sample of real homes, and “homes” in training are single rooms.

The authors list limitations themselves. Adding a new physical task still requires a human to write its prompt and author the keyframes of its motion. Vision alone gives no direct contact sensing, so contact-heavy tasks degrade. The orchestration layer is described as relatively simple. The paper reports rates for each layer but no single end-to-end latency figure.

For readers wanting to try it, the paper points to a project page at apple-aiml-research.github.io/ml-halo for code and videos. The page has a Code link, but we could not confirm from its content whether code is released or under what license. Until that is clear, HALO is a research result, not something anyone can download and run on a G1.