Apple Details Combined Training Method for Humanoid Robots
Apple has introduced an innovative training approach specifically designed for humanoid robots. This fresh methodology blends human guidance with robotic demonstrations, a technique Apple has dubbed “PH2D”.
This new training strategy was detailed shortly after Apple unveiled its advanced AI systems, Matrix3D and StreamBridge, illustrating the company’s continued focus on technological advancements. Apple’s history in robotics includes various projects, but this recent research zeroes in on humanoid robots.
The study, titled “Humanoid Policy ~ Human Policy,” critiques the limitations of conventional robot-training techniques and puts forward a more scalable and economically feasible alternative. Traditionally, robot training has heavily depended on robot demonstrators, a method criticized for being both labor-intensive and costly due to the need for teleoperated data collection.
Apple proposes a combined training method involving both human instructors and robot demonstrators. This innovative technique is not only meant to minimize training expenses but also to enhance the overall effectiveness of the training process. Employing consumer products, Apple identified a pathway to generate cost-effective training materials for humanoid robots.
In this context, the Apple Vision Pro was adapted by utilizing only the lower left camera for visual observation. Meanwhile, Apple’s ARKit was instrumental in capturing 3D head and hand movements. The system also employed a modified Meta Quest headset fitted with mini ZED Stereo cameras, providing a budget-friendly option for training purposes.
These customized headsets facilitated the training of humanoid robots in manipulating objects. Human instructors were recorded performing various hand movements, like grasping and lifting items or pouring liquids, while following verbal instructions. This documented footage was decelerated to serve as comprehensive training material for humanoid robots.
Apple developed a software model capable of integrating and processing training data from both humans and robots, known as “Physical Human-Humanoid Data” or PH2D. The processing unit responsible for this, termed the “Human-humanoid Action Transformer” or HAT, can adeptly handle inputs from both human and robotic sources.
This integrated approach allowed Apple’s researchers to create a unified policy framework drawing on both human and robot demonstration data. The result is a more generalized and robust training method, surpassing those relying solely on robot-generated data, according to the research.
The study underscores several advantages of this diversified training strategy. Not only does it offer a cost-effective solution, but robots trained using this methodology performed better on specific tasks than those trained exclusively with robot demonstrations. Task improvements were notably observed in operations like vertical object grasping.
Apple’s enhanced training methodology is expected to shape the development of upcoming products. While the robot-lamp prototype remains the only publicly demonstrated project to date, Apple is reportedly developing a mobile robotic assistant aimed at assisting with household chores and straightforward tasks.
As Apple continues to innovate in the realm of humanoid robotics, the company could revolutionize not only how robots are trained, but also how they interact with and assist humans in various environments. This approach potentially marks a pivotal shift in the dynamics of human-robot collaboration.
By pioneering methods that bridge the gap between human-led and robot-driven training, Apple is setting the stage for a future where humanoid robots can better comprehend and replicate human actions, enabling them to become more versatile and effective in everyday settings.
