Apple's HALO Lets a Unitree G1 Robot Navigate Unseen Homes

Trained in 1,000 virtual homes

The biggest obstacle in humanoid training is data, since people usually have to demonstrate behaviors through teleoperation or motion capture. Apple instead generated 1,000 virtual homes with different layouts and furniture, producing nearly 163,000 training episodes, about 300 hours of robot motion. Each example paired a command, the robot's camera view and a matching movement sequence. Invalid attempts, such as a foot passing through the floor, were filtered out automatically. Building the dataset took 48 hours on 250 Nvidia H100 GPUs.

Variety proved decisive. Trained in a single environment, the system succeeded 43% of the time when sitting in new simulated rooms. With 1,000 environments, that rose to 84%.

A Chinese robot with a Mac on its back

Apple did not build its own humanoid. The team used a Unitree G1, adding a head camera and a housing with a Mac mini M4 Pro with 64 GB of memory. The Mac interprets commands and plans movement, while the robot's Nvidia Jetson Orin NX handles the rest, so no external computer is needed.

HALO has three modules: one splits a command into simpler steps, another plans the next 5 seconds of whole-body motion about three times per second, and a 50 Hz controller drives the motors and keeps balance.

Where it still fails

Crouching under an obstacle succeeded in every trial, doorway passing in 93%, reaching an object in 92%, crawling in 90%, pushing boxes in 88% and sitting in 85%. Multi-step tasks fared worse, averaging 70% in a single real environment. The hardest sequence, crouching, approaching an object and pushing a box, worked only half the time. The researchers noted that after the first action the robot often ended up positioned so that the next target was out of view.