Apple Research Discloses HALO Autonomous Control System for Humanoid Robots Validated on Unitree G1
平均成功率91%!苹果研究团队披露机器人“大脑”进展：调用千问模型在宇树G1上验证
Apple Research disclosed the HALO autonomous control system for humanoid robots via an arXiv preprint titled 'Perceptive Humanoid Motion from Simulation.' The system was validated on a Unitree G1 robot in six unscanned real-world environments, achieving an average success rate of 91% across 360 single tasks. HALO employs a three-layer architecture: task orchestration using Qwen3.6-35B-A3B, motion planning with Qwen3.5-2B, and reinforcement learning-trained trackers for joint control. Training data was synthesized from 1,000 procedural simulation environments using 250 H100 GPUs over approximately 48 hours. The team notes that synthetic data generation still involves manual input for new physical tasks and that current performance relies on visual perception without tactile sensing.
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
Reported success rate
91%
Reported trials
View original evidence
图片来源：HALO项目页面截图拼接 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 S14Reported success rate
70%
Reported trials
trial_count
360
Reported trials
View original evidence
图片来源：HALO项目页面截图拼接 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 S14trial_count
10
Reported trials
Reported success rate
50%
Reported trials
training_data_hours
≈300 hours
View original evidence
研究团队在1000个程序生成的模拟家庭环境中，合成了约300小时的主要训练数据，将机器人“看到什么”“接到什么指令”和“如何运动”配对起来。
Open source S18gpu_count
250
synthesis_time_hours
≈48 hours
- paper: not_reported
- code: Reported pending Open source S5
- weights: not_reported
- dataset: not_reported
What remains unknown
Not established in the collected evidence: Environment, 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.
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
HALO的合成数据生成与训练流程：先生成动作轨迹，将其放入虚拟场景，再渲染机器人视角的画面，形成配套训练数据，用于训练动作生成器和动作跟踪器。
Open source S22
图片来源：HALO项目页面截图拼接 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 S14
另在一个真实环境中，对3种组合任务分别开展10次测试，平均成功率为70%。
Open source S15
其中，“俯身通过障碍—走向目标—推箱子”这一组合的成功率为50%。
Open source S16
研究团队在1000个程序生成的模拟家庭环境中，合成了约300小时的主要训练数据，将机器人“看到什么”“接到什么指令”和“如何运动”配对起来。
Open source S18
图片来源：HALO论文 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 S23
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 · msn.cn