AI News

Apple discloses the HALO system for humanoid robots, with an average task success rate of 91% in real environments

East MoneyUpdated Nov 9 08:46

AI summary

Apple's research team submitted the paper "Perceptive Humanoid Motion from Simulation" on arXiv, disclosing the humanoid robot autonomous control system HALO, which completed 360 single-task tests in 6 real environments that had not been pre-scanned or mapped, with an average success rate of 91%, and 3 combined tasks each tested 10 times with an average success rate of 70%. HALO is deployed on Unitree's G1 and is divided into three layers: the top layer uses Qwen3.6-35B-A3B running locally on a Mac mini carried by the robot for task orchestration; the middle layer uses Qwen3.5-2B as a frozen encoder together with a DiT decoder to plan full-body motions for about the next 5 seconds; and the bottom-layer tracker outputs joint control commands at a frequency of 50 times per second. The team pointed out that adding new physical tasks still requires manually writing prompts and designing keyframes, the robot mainly relies on visual perception, and performance declines when objects leave the field of view or are occluded; adding tactile perception is one of the future directions.

Why read this

AI score 72

The paper discloses the task success rate of HALO in real rooms and details of its three-layer architecture, and also points out current limitations such as the lack of tactile sensing.

Topics

Tags

Body · AI translation

Article body

On November 4 local time, Apple's research team submitted a paper titled "Perceptive Humanoid Motion from Simulation" to the preprint platform arXiv, disclosing progress on HALO, an autonomous control system for humanoid robots. In 360 individual task tests conducted in real environments, the average success rate reached 91%.

Apple discloses progress on robot "brain" research

According to the paper, Apple's research team deployed the HALO system on Unitree's G1 humanoid robot for validation, enabling the robot to carry the motions it learned in a simulated environment into unfamiliar real-world rooms for execution.

The research team conducted 360 individual task tests in 6 real environments that had not been pre-scanned or mapped, with an average success rate of 91%; in addition, they conducted 10 tests each for 3 combined tasks in one real environment, with an average success rate of 70%.

In terms of architecture, HALO is controlled collaboratively at three levels: the topmost layer is a task orchestration module based on a vision-language model, which by default uses Qwen3.6-35B-A3B and runs locally on a Mac mini carried by the robot, requiring no internet connection to work; the middle motion generation layer uses Qwen3.5-2B as a parameter-frozen encoder, which works with a DiT decoder to plan full-body motions for roughly the next 5 seconds based on the visual input and the current subtask; the bottom-layer motion tracker runs at a frequency of 50 times per second, converting motion plans into joint control commands.

The research team also pointed out current limitations: newly added physical tasks still require manually written prompts and designed motion keyframes; the robot mainly relies on visual perception and finds it difficult to directly perceive contact conditions, and its performance declines when objects leave the camera's field of view or are occluded. Adding tactile perception is one of the directions for subsequent improvement.

Humanoid robots refer to robots that imitate the human form and possess perception, decision-making, and motion capabilities; their "brain" is composed collaboratively of modules such as vision-language models, motion generation, and motion control.

CITIC Securities believes that with global giants entering the field, the physical AI industry is accelerating, and the robot-themed market conditions are now ripe for launch; advances by GPT-6 and Gemini 4 in robot planning, spatial perception, and Real-to-Sim modeling are broadening the space for general large models to participate in robot development and training, and Tesla is expected to release Optimus V3 in the fourth quarter, bringing the development of the humanoid robot industry into a new stage.

Guojin Securities pointed out that the core change in this round of Physical AI is that general models are beginning to move beyond perception and understanding into planning, decision-making, and action spaces, and the value of robot electronic hardware will spread from SoCs to storage, interconnect, and control segments; model training requires the advance accumulation of large amounts of real interaction data, and demand on the training side is expected to scale up before the robot bodies themselves do so on a large scale.

Guoyuan Securities stated that body manufacturers have entered a period of business validation, and commercialization paths are accelerating their divergence; demand for fine manipulation is further raising the configuration value of dexterous hands, driving coordinated upgrades in micro-drives, precision transmission, and tactile perception.

Multiple stocks actively snapped up by margin traders this month

According to the concept sectors on East Money, there are currently more than 300 stocks in the A-share market related to the humanoid robot concept, with a combined total market value of about 6.55 trillion yuan. Midea Group, Unitree, AVIC Chengdu Aircraft, Inovance Technology, and others rank among the largest.

Since the beginning of the year, the humanoid robot concept sector has generally trended weakly, but since November it has strengthened significantly: Yongzhen shares surged 49% to rank first, while Quectel, Huashi Technology, Zhenyu Technology, MediTech, Awinic, and many other stocks all rose by more than 20%.

In terms of performance, a total of 267 concept stocks achieved profits in the first three quarters, of which 156 saw positive year-on-year growth in net profit. Among them, Midea Group ranked first with a net profit of 41.393 billion yuan, GigaDevice's net profit reached 11.066 billion yuan, a year-on-year increase of more than 9 times, and Inovance Technology and Huaqin Technology each had net profits exceeding 4.3 billion yuan. At the same time, 65 stocks including GAC Group, Lianchuang Electronics, and OFILM recorded losses in the first three quarters.

From a capital perspective, East Money Choice data shows that since November, a total of 30 humanoid robot concept stocks have obtained net margin financing purchases exceeding 10 million yuan.

Among them, Rockchip received 252 million yuan in additional leverage funds, Huashi Technology and Everwin Precision were respectively snapped up for 153 million and 133 million yuan, and Dayang Electric, Tianyu Digital, Unitree, and SWANCOR all received net margin financing purchases exceeding 80 million yuan.

Rockchip's net profit attributable to shareholders in the first three quarters was 1.332 billion yuan, up 70.90% year on year. The company's RK3588+RK1828 dual-chip platform can deploy vision-language models on the device side, supporting environmental understanding and task planning; new products such as RK3572 have already entered mass production in the third quarter, the RK182X coprocessor is expected to achieve volume production in the fourth quarter, and the next-generation on-device AI coprocessor RK1860 is planned for launch in the fourth quarter.

Keli Sensing has developed six-axis force/torque sensors and joint torque sensors for humanoid robot scenarios, with product capabilities covering key parts such as wrists, ankles, and joints, and can provide high-precision force feedback and compliant control. Cumulative sales of the company's robot force sensors exceeded 2,000 units in the first half of the year, with monthly orders continuing to exceed 1,000 units.

(Article source: East Money Research Center) Middle of the article Embedded advertisement

Source

East Money