White humanoid robot carrying a cardboard box through a dusty residential attic
AI-generated image

Research

Apple researchers train a Unitree G1 to move through unseen rooms using mostly simulated data

A team of Apple researchers has trained a Unitree G1 humanoid to walk to objects, sit on chairs, crawl under tables and push boxes in rooms it had never seen, learning mostly from motion data generated in simulation. Humanoids Daily detailed the system, called HALO, on 15 November; the paper was posted to arXiv earlier this month, and the code is on Apple's research GitHub.

The robot completed 91% of 360 single-task trials across six real, unmapped environments without any teleoperation data, Humanoids Daily reports. Chaining skills was harder, with an average of 70% across three sequences, and some chains, such as crouching under a table and then sitting, worked in only 60% of tries in a single test environment.

The training data came from a pipeline that generates human-like motions with a diffusion model, places them in 1,000 procedurally generated rooms and renders them in NVIDIA's Isaac Sim from the robot's own camera viewpoint, producing about 296 hours of footage. Some human data remains, including public motion-capture clips used to train the motion tracker.

On the robot, an off-the-shelf vision-language model breaks a request such as "go sit on the chair" into steps, a motion generator plans a few seconds of whole-body movement from the camera image, and a reinforcement-learning tracker follows that plan at 50 Hz. The setup adds a wide-angle camera and a Mac mini on the robot's back, and runs untethered.

The authors list limits: each new skill still needs a person to write prompts and keyframes, the robot has no sense of touch, and none of the six skills involves grasping objects. Humanoids Daily notes the paper does not compare HALO with other published systems under matching conditions.

More on: Unitree · China · Research · Pilots and deployments

← All news