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, validated on Unitree G1 hardware in six unscanned real-world environments with an average single-task success rate of 91%. The system employs a three-layer architecture: a local visual language model (Qwen3.6-35B-A3B) for task orchestration, a Qwen3.5-2B model for action generation, and a reinforcement learning-based tracker for joint control. Training data was synthesized from 1,000 simulated environments over approximately 300 hours using 250 NVIDIA H100 GPUs.
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.
- 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
Reported success rate
50%
Reported trials
training_hours
≈48 hours
simulated_environment_count
1,000
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 S18training_data_duration
≈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 S18trial_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 S14gpu_count
250
- 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
具体而言，系统先生成动作，再将动作放进虚拟房间，最后从机器人的视角生成画面。
Open source S19
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
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
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 — Hardware and sensing · thepaper.cn