May 22 2025 – Yesterday, AppleInsider, a well-known tech media outlet, reported that Apple Inc. has recently unveiled a novel approach called PH2D (Physical Human-Humanoid Data) for training humanoid robots. This innovative method integrates human coaches with robot demonstrators to achieve more efficient learning outcomes for the robots.
On Wednesday this week, Apple published a research paper titled “Humanoid Policy ~ Human Policy,” highlighting the shortcomings of traditional training methods and introducing a scalable and cost-effective solution. The paper emphasizes that traditional training for humanoid robots often relies heavily on robot demonstrators, a process described as “labor-intensive” and requiring expensive remote operation data collection.
According to the blog post from AppleInsider, Apple’s new approach combines human coaches and robot demonstrators, utilizing modified consumer devices to create training materials. For instance, the Apple Vision Pro headset has been adapted to use only the bottom-left camera for visual observation, paired with ARKit technology to capture 3D head and hand pose data. Similarly, a modified Meta Quest headset, equipped with a mini ZED stereo camera, serves as a low-cost training tool.
Human coaches use these devices to record hand movements, including actions like grasping, lifting objects, and pouring liquids. During the recording process, voice guidance is also provided. The recorded videos are then slowed down for robot training purposes.
To further enhance the training process, Apple has developed a model named “Human-humanoid Action Transformer” (HAT). This model is capable of processing data generated by both human coaches and robot demonstrators simultaneously. By constructing a universal policy framework, the HAT model demonstrates superior generalization capabilities and robustness compared to training methods that rely solely on real robot data. Research indicates that this combined training strategy outperforms single-robot demonstrator training in specific tasks, such as vertical object grasping.
Apple’s research suggests that this integrated training approach not only offers cost-effectiveness but also significantly improves robot performance. Although only a robot lamp prototype has been showcased so far, rumors suggest that Apple is actively developing mobile robots for end consumers, capable of performing household chores and simple tasks.
