Apple Unveils HALO Humanoid Robot Research Using Unitree G1 Body and Mac mini M4 Pro
苹果罕见公布人形机器人进展，宇树 G1 背着一台跑千问的 Mac mini
Apple's research team has published details on the HALO project, a humanoid robot system combining a Unitree G1 body with a Mac mini M4 Pro running Alibaba's Qwen models. The system uses synthetic data generated via simulation to train for real-world tasks like sitting and navigating unknown environments. In physical tests across six unseen homes, it achieved an average success rate of 91% on individual 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
simulation hours
300 hours
Reported duration
9 hours
View original evidence
论文里还提到，仅生成 SIT（坐下）一个技能的 5 万条样本，就用了 250 张 H100 GPU 跑了 9 个小时；全部 1178 个房间的生成用了约 300 张 GPU、26 个小时。
Open source S23Reported duration
26 hours
View original evidence
论文里还提到，仅生成 SIT（坐下）一个技能的 5 万条样本，就用了 250 张 H100 GPU 跑了 9 个小时；全部 1178 个房间的生成用了约 300 张 GPU、26 个小时。
Open source S23simulation rooms
1,000
View original evidence
研究团队用 运动扩散模型 生成全身动作，用程序化生成器（ INFINIGEN ）造出 1000 个虚拟家庭，客厅、办公室、卧室，家具疏密随机；再在 Isaac Sim 里把动作放进场景，渲染出机器人头部相机看到的第一视角画面，最后交给自动校验器筛掉穿模、滑脚的坏样本。
Open source S19simulation rooms
1,178
View original evidence
论文里还提到，仅生成 SIT（坐下）一个技能的 5 万条样本，就用了 250 张 H100 GPU 跑了 9 个小时；全部 1178 个房间的生成用了约 300 张 GPU、26 个小时。
Open source S23
Dataset details reported by the source
- Tasks
- sit Open source S23
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
What remains unknown
Dataset licensing and commercial-use rights are not established here.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
团队在约 1000 个程序生成的房间里做出了 16.3 万条、近 300 小时 的配对样本，数据规模只受算力限制。
Open source S22
一条常见的数据样本里，既要有它听到的指令和第一视角画面，也要有完成任务时全身关节怎么动的记录。
Open source S14
论文里还提到，仅生成 SIT（坐下）一个技能的 5 万条样本，就用了 250 张 H100 GPU 跑了 9 个小时；全部 1178 个房间的生成用了约 300 张 GPU、26 个小时。
Open source S23
最近，苹果研究团队公开了一项名为 HALO 的人形机器人研究，直接用宇树 G1 当身体，后背绑着一台 Mac mini M4 Pro，里面运行的还是阿里的千问模型。
Open source S6
动作生成模型据此给出多组轨迹 ，程序检查脚底有没有打滑、身体有没有穿过地面、关节有没有超过活动范围，再把筛过的动作和椅子摆进虚拟房间，避开周围的家具。
Open source S20
研究团队用 运动扩散模型 生成全身动作，用程序化生成器（ INFINIGEN ）造出 1000 个虚拟家庭，客厅、办公室、卧室，家具疏密随机；再在 Isaac Sim 里把动作放进场景，渲染出机器人头部相机看到的第一视角画面，最后交给自动校验器筛掉穿模、滑脚的坏样本。
Open source S19
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：Chinese robotics — Hardware and sensing · news.qq.com