Skip to content
Chinese robotics — Research news discovery·· 3 days agoSignalEditorial score75

Apple Research Team Discloses Progress on Humanoid Robot 'Brain' HALO

苹果披露机器人“大脑” 研究进展

Summary

Apple's research team submitted a preprint titled 'Perceptive Humanoid Motion from Simulation' to arXiv, detailing the HALO autonomous control system validated on Unitree G1 robots. The system transfers actions learned in simulation to real-world environments using a three-layer architecture involving visual language models for task orchestration and reinforcement learning for joint control. The team notes that simulation data generation still required manual prompt design and human motion capture, and the system currently lacks direct tactile perception.

Source: Chinese robotics — Research news discovery · Read original article ↗

English translation unavailable. Choose the original article to read it.

What the source reports

Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.

Control
Reported autonomousOpen source S3

Reported numbers

  • GPU hours for data synthesis

    ≈48 hours

    View original evidence
    图片来源:HALO项目页面截图拼接 据介绍,研究团队使用250张英伟达H100 GPU,耗时约48小时合成主要训练数据。
    Open source S18
  • Training data synthesis duration

    ≈48 hours

    View original evidence
    图片来源:HALO项目页面截图拼接 据介绍,研究团队使用250张英伟达H100 GPU,耗时约48小时合成主要训练数据。
    Open source S18
  • Action planning frequency

    ≈3

    View original evidence
    这一层每秒重新规划约3次,使动作计划能够随环境变化持续更新。
    Open source S14
  • Action tracker execution frequency

    50

    View original evidence
    底层则是经强化学习、策略蒸馏等步骤训练的动作跟踪器,以每秒50次的频率运行,结合深度图像和机器人自身状态,将动作计划转化为关节控制指令。
    Open source S15
  • paper: not_reported
  • code: Reported pending Open source S5
  • weights: not_reported
  • dataset: not_reported
  • code: not_reported
Source excerpts and review record

Automatically extracted; no manual editorial approval recorded.

研究团队将该系统部署在 宇树科技 的G1人形机器人上进行验证,让机器人将在仿真环境中学到的动作,带到陌生的现实房间中执行。

Open source S4

当地时间11月4日,苹果研究团队在预印本平台arXiv提交论文《Perceptive Humanoid Motion from Simulation》,披露了人形机器人自主控制系统HALO。

Open source S3

图片来源:HALO项目页面截图拼接 据介绍,研究团队使用250张英伟达H100 GPU,耗时约48小时合成主要训练数据。

Open source S18

动作生成器通过模仿学习训练,使用64张B200;动作跟踪器经强化学习、策略蒸馏等步骤训练,使用4张H100。

Open source S19

这一层每秒重新规划约3次,使动作计划能够随环境变化持续更新。

Open source S14

底层则是经强化学习、策略蒸馏等步骤训练的动作跟踪器,以每秒50次的频率运行,结合深度图像和机器人自身状态,将动作计划转化为关节控制指令。

Open source S15

研究团队还公开了数据生成、模型训练和部署代码,涵盖动作生成、动作跟踪、任务编排以及仿真和真实机器人运行等环节。

Open source S5

Source:Chinese robotics — Research news discovery · finance.sina.com.cn

Timezone · UTC

Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).