Apple's HALO trains a Unitree G1 in 1,000 simulated homes and hits 91% in six real ones it never saw

A team of Apple researchers has trained a humanoid robot to sit on chairs, walk to named objects, step through half-open doors, crouch and crawl under tables and push boxes with no teleoperation and no scan of the homes it works in. In a paper posted to arXiv on Nov. 4, they report that the system, called HALO, succeeded in 91% of 360 trials on a Unitree G1 in six real homes it had never seen, scanned or mapped. It learned those skills from data generated in 1,000 procedurally generated rooms in simulation.

The paper sets HALO against teleoperation and capture suits, where an hour of training data costs at least an hour of human time. Apple instead spent compute: synthesizing the main corpus, about 300 hours of episodes, took about 48 hours on 250 Nvidia H100 GPUs, the paper says. That trade works for moving through a home. The paper's own numbers also show that it shifts the bottleneck instead of removing it.

More rooms stopped helping after 100

In the ablation on the sit skill, scored in held-out simulated rooms, training on one room gave 43% success, 10 rooms 64% and 100 rooms 81%. Going to 1,000 rooms added 3 more points, to 84%. Most of the gain from scene variety came early.

Chaining skills is where the system still struggles. Across three multi-step tasks in one real room, success averaged 70%. "Crouch, walk to the target, push the box" worked in half of its 10 runs. In 3 of the 5 failures, the crouch ended with the robot's head turned away from the box, and it never found it again. In two Crouch→Sit runs, it tried to sit on the sofa instead of the chair.

A person still writes every new skill

The pipeline removes the operator from data collection, "but not from task design," the authors write. Each new task needs a human to write its prompt and set the keyframes of its motion. The crawl and crouch motions started from phone video, and pushing used motion capture. The robot also relies on vision and has no direct sense of contact, so contact-heavy tasks degrade once an object leaves the camera's view. The authors name tactile sensing as the natural fix.

The hardware is ordinary by design: the G1 carries an extra head camera and a Mac mini M4 Pro in a rear enclosure, and the whole stack, including a language model that breaks requests into subtasks, can run onboard without an internet connection.

A day later, Shanghai's Archon Robotics went the other way and released a controller that makes teleoperation on the same G1 nearly five times as productive. HALO's six skills are mostly about getting around a room. Archon's test tasks, kicking a box and picking up slippers, are the contact-heavy kind that HALO's authors say vision alone handles poorly. For now, simulation is replacing the operator in walking and sitting, and the operator still owns the work done with hands and feet.