Apple’s HALO humanoid robot control system hits 91% average task success in 360 tests

Cointime update: Apple’s research team published a paper on arXiv titled “Perceptive Humanoid Motion from Simulation,” disclosing HALO, an autonomous control system for humanoid robots. In 360 single-task tests across six previously unscanned real environments, HALO achieved a 91% average success rate; across three combined-task scenarios tested 10 times each, the average was 70%, with the “duck under obstacle—walk to target—push box” sequence succeeding 50% of the time. The team used 250 Nvidia H100 GPUs for about 48 hours to synthesize main training data.