Apple research team releases humanoid robot control system HALO, based on Unitree G1 and Qwen model
Apple's research team submitted the paper "Perceptive Humanoid Motion from Simulation" to arXiv, disclosing the autonomous humanoid robot control system HALO. The hardware uses Unitree Robotics' G1 with a built-in Mac mini (M4 Pro chip, 64GB unified memory) mounted on its back. HALO has three layers of control: the task orchestrator runs the Qwen large model Qwen3.6-35B-A3B by default, updating approximately every 2 seconds; the motion generator reads first-person RGB images at about 3Hz, generating full-body motion plans for the next 5 seconds; the motion tracker sends commands to joint actuators 50 times per second. The team conducted 360 single-task tests in 6 real environments that were not pre-scanned, achieving an average success rate of 91%, and an average success rate of 70% across 3 combinations of long-sequence composite tasks.
Why read this
AI score 67
The original text discloses HALO's three-layer control structure and real-environment test success rates, offering insight into one engineering approach to autonomous humanoid robot control.
Topics
Body · AI translation
Article body
On Monday, robot concept stocks strengthened, with Xiangming Intelligent, Shangwei New Materials and others hitting the 20CM limit-up, Unitree Robotics rising more than 9%, and Tianzhun Technology, Kaite Shares and many other stocks following the gains.
On the news front, according to The Paper, Apple's research team disclosed progress on research into a robot "brain." On November 4 local time, Apple's research team submitted a paper titled "Perceptive Humanoid Motion from Simulation" on the preprint platform arXiv, disclosing the humanoid robot autonomous control system HALO.
The research team used the Unitree G1 humanoid robot, with a dedicated 3D-printed protective shell mounted on its back, housing a Mac mini (M4 Pro chip, 64GB unified memory).
When executing tasks, HALO is controlled collaboratively by three layers. The top layer is the task orchestrator, which by default uses the Qwen large model Qwen3.6-35B-A3B; the system can also choose to call a cloud model via API, using a vision-language model to understand natural language instructions and break open-ended requests such as "go to the living room and get the box" into a series of subtasks; this module runs by default on the onboard Mac mini and updates about once every 2 seconds.
The middle layer is the motion generator, which reads first-person RGB images and the current subtask at a frequency of about 3Hz and generates a full-body motion plan for the next 5 seconds, including joint angles and pelvis pose.
The bottom layer is the motion tracker, which runs at a frequency of 50 times per second, combines depth information and proprioception, sends control commands to each joint actuator, and adapts to balance, contact and actuator constraints through reinforcement learning.
According to the paper, the research team conducted 360 single-task tests in 6 real environments that had not been pre-scanned or mapped, with an average success rate of 91%; for long-sequence composite tasks, in a single real scene, 10 tests were conducted for each, and the average success rate for 3 combinations was 70%.
HALO's training data all comes from procedurally generated simulated home environments. The Apple team used about 1,000 procedurally generated rooms to synthesize about 300 hours of paired video, language and motion data, covering scenes such as living rooms, offices and bedrooms; in the data generation stage, no personnel need to enter real residences to scan the environment, nor is continuous human teleoperation required.
The paper states that data synthesis took about 48 hours on 250 H100 GPUs, training the motion generator used 64 B200 GPUs, and training the motion tracker used 4 H100 GPUs.
A Zhongtai Securities research report pointed out that the brain is one of the core variables for embodied intelligence to move from demonstration to practicality. Many robot manufacturers have publicly listed insufficient generalization capability as the biggest constraint, and technical routes are also competing around the integration of VLA and world models. Capital, policy and industrial strategy are concentrating on the brain in the same time window, reflecting the industry's recognition of the judgment that "the threshold for the body is falling, and the brain determines the ceiling."
According to the Zhongtai Securities research report, on the capital side, funds are concentrating toward embodied brains. In the first half of 2026, domestic embodied intelligence track financing was about 43.8 billion yuan, of which companies with embodied brains as their core business absorbed 22.253 billion yuan, accounting for 50.8%, while robot body and complete machine companies raised only 5.6 billion yuan in the same period, accounting for 12.8%.
For example, Unitree Robotics' previous IPO plan was to raise 4.202 billion yuan, of which 2.022 billion yuan was to be invested in intelligent robot model R&D projects, focusing on the two major directions of WMA world model-action and VLA vision-language-action.
(Source: East Money Research Center) Middle of the article Embedded advertisement