Apple researchers submitted a paper to arXiv on Nov. 4, Perceptive Humanoid
Motion from Simulation, detailing HALO, a humanoid autonomous-control system.
They deployed HALO on Unitree’s G1 humanoid to transfer simulation-trained
motions into unfamiliar real rooms and published data, training and deployment
code covering motion generation, tracking, task sequencing and sim-to-real
operation. In six unscanned real environments the team ran 360 single-task
trials with an average success rate of 91%. In one real environment they ran
three composite tasks 10 times each, averaging 70% success; the sequence
duck-under-obstacle → walk-to-target → push-box succeeded 50% of the time.