Apple research team revealed progress in robot “brain” research. On November 4, local time, the Apple research team submitted the paper “Perceptive Humanoid Motion from Simulation” on the preprint platform arXiv, revealing HALO, an autonomous control system for humanoid robots. The research team also disclosed data generation, model training, and deployment codes, covering motion generation, motion tracking, task orchestration, and simulation and actual robot operation. According to the paper, the research team conducted 360 single-task tests in 6 real environments without prior scanning or mapping, with an average success rate of 91%. Also, in a real environment, each of the 3 combined tasks was tested 10 times, with an average success rate of 70%. Among them, the success rate of the combination “lean over the obstacle - towards the goal - push the box” was 50%. According to reports, the research team used 250 Nvidia H100 GPUs and it took about 48 hours to synthesize the main training data. The motion generator uses 64 sheets of B200 by imitating learning and training; the motion tracker uses 4 H100 sheets after training through steps such as reinforcement learning and strategic distillation. It is important to note that simulation-generated data is not completely free from human participation. The team said that the new physical tasks still require manual writing of prompts and design action keyframes, and some actions also use human video or motion capture material. At the same time, the research team pointed out that currently robots mainly rely on visual perception, making it difficult to directly sense contact situations; when objects leave the camera's field of view or are blocked, the performance of tasks involving more contact will decrease. Adding tactile perception is one of the future directions of improvement.
Zhitongcaijing · 2d ago
Share
Listen to the news
Apple research team revealed progress in robot “brain” research. On November 4, local time, the Apple research team submitted the paper “Perceptive Humanoid Motion from Simulation” on the preprint platform arXiv, revealing HALO, an autonomous control system for humanoid robots. The research team also disclosed data generation, model training, and deployment codes, covering motion generation, motion tracking, task orchestration, and simulation and actual robot operation. According to the paper, the research team conducted 360 single-task tests in 6 real environments without prior scanning or mapping, with an average success rate of 91%. Also, in a real environment, each of the 3 combined tasks was tested 10 times, with an average success rate of 70%. Among them, the success rate of the combination “lean over the obstacle - towards the goal - push the box” was 50%. According to reports, the research team used 250 Nvidia H100 GPUs and it took about 48 hours to synthesize the main training data. The motion generator uses 64 sheets of B200 by imitating learning and training; the motion tracker uses 4 H100 sheets after training through steps such as reinforcement learning and strategic distillation. It is important to note that simulation-generated data is not completely free from human participation. The team said that the new physical tasks still require manual writing of prompts and design action keyframes, and some actions also use human video or motion capture material. At the same time, the research team pointed out that currently robots mainly rely on visual perception, making it difficult to directly sense contact situations; when objects leave the camera's field of view or are blocked, the performance of tasks involving more contact will decrease. Adding tactile perception is one of the future directions of improvement.
Disclaimer:Webull uses external vendor Google Translation Service for news translations where we endeavour to ensure these are correct, however, we recommend that you please double-check this information accordingly. Webull is not responsible for translation errors or issues.
What's Trending