Skip to content
Chinese robotics — Hardware and sensing·· 6 days agoSignalEditorial score75

Apple Research Team Unveils HALO Autonomous Control System for Humanoid Robots

苹果公开人形机器人“大脑”HALO!搭载宇树G1测试 陌生环境任务成功率达91%

Summary

Apple's research team has published findings on HALO, an autonomous control system validated on Unitree G1 hardware. The system combines visual language models with motion generation to perform tasks like pushing boxes and navigating obstacles in unscanned environments. While single tasks achieved a 91% average success rate across six real-world settings, composite tasks averaged 70%. The architecture relies on local Qwen models and synthetic data from 1,000 simulated homes, though it still requires manual prompt design for new physical tasks.

Source: Chinese robotics — Hardware and sensing · 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.

Reported numbers

  • Reported success rate

    91%

    Reported trials

    View original evidence
    6种陌生环境测试成功率91% 复合任务仍有改善空间 在实际测试方面,研究团队选择6个未经事先扫描或建图的真实环境,进行360次单项任务测试,平均成功率达91%。
    Open source S16
  • Reported success rate

    70%

    Reported trials

    View original evidence
    研究团队另外在单一真实环境中,针对3种复合任务各进行10次测试,平均成功率为70%。
    Open source S18
  • Reported success rate

    50%

    Reported trials

    View original evidence
    其中,“弯腰通过障碍物、走向目标、推动箱子”这项连续任务的成功率仅50%,显示机器人处理多步骤动作时仍面临挑战。
    Open source S19
  • training data duration

    ≈300 hours

    View original evidence
    为了训练机器人适应不同环境,研究团队建立一套大规模合成资料生成流程,在1,000个程序生成的模拟家庭环境中,制作约300小时的主要训练资料,将机器人看到的画面、接收的指令及对应动作配对。
    Open source S20
  • GPU hours

    ≈48 hours

    View original evidence
    在计算资源方面,研究团队使用250张英伟达H100 GPU,耗时约48小时完成主要训练资料的合成;动作生成器采用模仿学习训练,使用64张B200 GPU;动作追踪器则通过强化学习、策略蒸馏等方式训练,使用4张H100 GPU。
    Open source S23
  • simulated environments

    1,000

    View original evidence
    为了训练机器人适应不同环境,研究团队建立一套大规模合成资料生成流程,在1,000个程序生成的模拟家庭环境中,制作约300小时的主要训练资料,将机器人看到的画面、接收的指令及对应动作配对。
    Open source S20
  • action replanning frequency

    ≈3

    View original evidence
    系统每秒重新规划约3次,让机器人能随着环境变化持续修正动作。
    Open source S14
  • paper: not_reported
  • code: not_reported
  • weights: not_reported
  • dataset: not_reported

What remains unknown

Not established in the collected evidence: Environment, Control, Data origin.

Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.

Source excerpts and review record

Automatically extracted; no manual editorial approval recorded.

6种陌生环境测试成功率91% 复合任务仍有改善空间 在实际测试方面,研究团队选择6个未经事先扫描或建图的真实环境,进行360次单项任务测试,平均成功率达91%。

Open source S16

研究团队另外在单一真实环境中,针对3种复合任务各进行10次测试,平均成功率为70%。

Open source S18

其中,“弯腰通过障碍物、走向目标、推动箱子”这项连续任务的成功率仅50%,显示机器人处理多步骤动作时仍面临挑战。

Open source S19

为了训练机器人适应不同环境,研究团队建立一套大规模合成资料生成流程,在1,000个程序生成的模拟家庭环境中,制作约300小时的主要训练资料,将机器人看到的画面、接收的指令及对应动作配对。

Open source S20

在计算资源方面,研究团队使用250张英伟达H100 GPU,耗时约48小时完成主要训练资料的合成;动作生成器采用模仿学习训练,使用64张B200 GPU;动作追踪器则通过强化学习、策略蒸馏等方式训练,使用4张H100 GPU。

Open source S23

系统每秒重新规划约3次,让机器人能随着环境变化持续修正动作。

Open source S14

Source:Chinese robotics — Hardware and sensing · c114.net.cn

Timezone · UTC

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