Humanoid single-task success rate 91%... consecutive tasks also 70%

Apple researchers unveiled HALO, a humanoid autonomous control system that achieved a 91% average success rate across 360 single-task trials in six real indoor spaces that were neither pre-scanned nor mapped, with multi-step tasks chaining three motions succeeding 70% of the time. The system was tested on a Unitree G1 robot fitted with added cameras and compute, per a paper pre-published on arXiv on the 4th. The work targets the scarcity of paired instruction, video, and whole-body motion data needed to translate spoken commands and camera input into full-body actions. The results matter as a sign that learned humanoid control can generalize to unseen, unmapped environments rather than relying on pre-built maps. Next, watch whether follow-up work scales beyond six sites and three-motion sequences toward longer tasks and broader real-world deployment.