Apple Sent a Unitree Robot Into Six Homes It Had Never Seen. It Finished 91 Percent of the Tasks

Apple's robotics work has been quiet for a year, and then arrives with a number worth arguing about. In a paper posted to arXiv on Nov. 4, its machine learning team described HALO -- Humanoid Autonomy, Learned in simulation, running Onboard -- a three-layer control stack trained without a single hour of teleoperated robot footage and deployed on a Unitree G1. Across 360 single-task trials in six real homes the system had never scanned or mapped, it succeeded 91 percent of the time. The code for data generation, training and deployment was released alongside the paper.

The data engine is the actual contribution. Apple's pipeline procedurally generated 1,200 residential interiors, kept 1,178 that passed validity checks, and split them into 1,000 for training and 178 for testing, producing 296.3 hours of paired synthetic data and roughly 163,000 training clips on 250 H100 GPUs in about 48 hours, with no human in the loop during generation. The comparison the authors draw is blunt: an hour of teleoperated data costs about an hour of human labour, and the resulting policy tends to break when the room changes.

The stack runs on the robot. A vision-language model, Qwen3.6-35B-A3B by default, sits on a tethered Mac mini, parses instructions, decomposes them into subtasks and decides when to retry. A mid-level module pairs a Qwen3.5-2B encoder with a diffusion-transformer decoder to emit whole-body trajectories about five seconds long, and a reinforcement-learned tracker at the bottom keeps the machine balanced while it executes. Cloud models can be called for extra reasoning, but the default loop is local.

Single tasks -- walk to an object, sit on a sofa, sidle through a half-open door, duck under a desk, push a box -- landed at 91 percent. Composition is where it thins out: three multi-step tasks averaged 70 percent, and the long sequence of ducking an obstacle, crossing the room and pushing a box succeeded half the time. That gap between one clean instruction and a chain of them is the same wall every humanoid lab is currently hitting.

The limitations are stated plainly. HALO is vision-only, so contact-heavy work degrades when an object leaves the camera's view or is occluded, and tactile sensing is listed as the next step. New tasks still require a human to write prompts and design keyframes, and some motions were seeded from human video or motion-capture reference, so the claim is automated scene generation rather than automated skill discovery. The authors also note the six test homes were empty of residents during trials.

What Apple has shown is narrower and more useful than a robot demo: the bottleneck in humanoid manipulation is shifting from hardware and model architecture to the supply of training data, and simulation can now manufacture that supply at a scale teleoperation cannot match. Whether a policy trained on synthetic rooms survives a lived-in kitchen with a wet floor, a pet and a chair in the wrong place is the question the paper does not answer. Apple has published no product timeline and this is research, not a roadmap. But it is the first time the company has put a number on how far its robot brain gets without a human holding the controller.