Apple’s HALO humanoid-control system achieved 91% average success across 360 single-task tests in six previously unmapped real environments.

Apple researchers disclosed HALO, an autonomous humanoid-control system, in the preprint “Perceptive Humanoid Motion from Simulation,” submitted to arXiv on November 4. They evaluated the system on a Unitree G1 humanoid robot, transferring motions learned in simulation into unfamiliar physical rooms.

The reported 91% average success rate covers 360 single-task tests in six real environments that had not been scanned or mapped beforehand. Demonstrations included pushing boxes, following people, sitting on chairs, moving sideways through partially open doors, and bending or crawling beneath tables.

More complex sequences were evaluated separately in one real environment, with 10 tests for each of three combined tasks and an average success rate of 70%. The sequence “bend through an obstacle, walk toward a target, push a box” achieved 50% success, so the single-task result does not describe combined-task performance.

The researchers reported that the robot mainly relies on visual perception and has difficulty directly sensing contact. Performance on contact-intensive tasks deteriorated when objects left the camera’s field of view or became occluded; tactile sensing was identified as a possible improvement direction, not a demonstrated capability.