Apple Discloses HALO Humanoid Robot System Validated on Unitree G1 with Qwen Models
苹果披露人形机器人 HALO 系统：调用千问模型在宇树 G1 上验证，单项任务平均成功率 91%
Apple's research team disclosed the HALO autonomous control system for humanoid robots via an arXiv paper. The system, validated on a Unitree G1 robot equipped with a Mac mini (M4 Pro), utilizes Alibaba's Qwen models for task orchestration and motion generation. Testing across six unscanned environments yielded an average single-task success rate of 91%. The architecture combines visual-language models, diffusion-based motion generation, and reinforcement learning policies, though current performance relies on visual perception and requires manual intervention for certain physical tasks.
Source: IT Home — Robotics · 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
Reported success rate
91%
Reported trials
Reported success rate
70%
Reported trials
View original evidence
针对长序列复合任务，在单一真实场景，各进行 10 次测试，3 种组合平均成功率为 70%；其中最复杂的“俯身通过障碍 — 走向目标 — 推箱子”10 次测试成功 5 次，成功率为 50%。
Open source S9Reported success rate
50%
Reported trials
View original evidence
针对长序列复合任务，在单一真实场景，各进行 10 次测试，3 种组合平均成功率为 70%；其中最复杂的“俯身通过障碍 — 走向目标 — 推箱子”10 次测试成功 5 次，成功率为 50%。
Open source S9test_count
360
Reported trials
test_count
10
Reported trials
View original evidence
针对长序列复合任务，在单一真实场景，各进行 10 次测试，3 种组合平均成功率为 70%；其中最复杂的“俯身通过障碍 — 走向目标 — 推箱子”10 次测试成功 5 次，成功率为 50%。
Open source S9training_hours
≈48 hours
Dataset details reported by the source
- Embodiments
- humanoid Open source S3
- paper: Reported pending Open source S3
- dataset: not_reported
What remains unknown
Not established in the collected evidence: Environment, Data origin.
Dataset licensing and commercial-use rights are not established here.
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.
据论文披露， 研究团队在 6 个未预先扫描或建图的真实环境中开展 360 次单项任务测试，平均成功率为 91% 。
Open source S8
IT之家 11 月 7 日消息，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
不过，仿真生成数据并未完全摆脱人工参与。
Open source S12
研究团队同时指出，目前机器人主要依赖视觉感知，难以直接感知接触情况；当物体离开相机视野或被遮挡时，涉及较多接触的任务表现会下降。
Open source S14
加入触觉感知将是后续改进方向之一。
Open source S15
针对长序列复合任务，在单一真实场景，各进行 10 次测试，3 种组合平均成功率为 70%；其中最复杂的“俯身通过障碍 — 走向目标 — 推箱子”10 次测试成功 5 次，成功率为 50%。
Open source S9
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 S10
IT之家附相关链接： 论文： https://arxiv.org/abs/2611.06908 项目主页： https://apple-aiml-research.github.io/ml-halo/ GitHub 仓库： https://github.com/apple-aiml-research/ml-halo 广告声明：文内含有的对外跳转链接（包括不限于超链接、二维码、口令等形式），用于传递更多信息，节省甄选时间，结果仅供参考，IT之家包含外链的文章均包含本声明。
Open source S16
Implications for data suppliers
RoboSignal interpretation and collection questions, not statements of buyer demand.
- Confirm the required data type and collection setting with the buyer; this source does not establish a complete collection specification.
- Compare the reported units and scope before using these quantities in a budget. Recording hours, sensor-hours and trajectories are different measures.
- Check the exact dataset version, license, modalities and task coverage against the intended use.
- Validate demand and acceptance criteria with a buyer before scaling. Publication, popularity and a research result do not establish a purchase commitment.
Source：IT Home — Robotics · ithome.com