Unitree: Humanoid Hype vs. Robotic Reality
Is China surging ahead in humanoid robots?
Contents
Things that caught our attention
We have shared some interesting bits and bops in our recent notes:
JD.com reincarnates it’s Joybuy domain as a B2C platform in the UK
Meituan’s internationalisation of food and grocery delivery (Keeta)
The Rise of "Human-Like" AI Agents: Why Monica's MANUS Deserves Your Attention
Chery’s IPO Sprint: A 28-Year Journey of Rise, Fall, and Revival
Someone is coming for your lunch (a word about Chinese internet companies going abroad)
The Gap Between the U.S. & China in AI Is Widening, Not Narrowing
Introduction
Humanoid robots are at a turning point. Machines that walk, talk, and integrate into the workforce have captured public imagination and investor interest alike. The pace of progress has been striking—2024 marked a major inflection point, driven by rapid advances in AI and automation. Yet, while technology is evolving faster than expected, fundamental challenges remain. The laws of physics, supply constraints, and economic realities still apply. Even the most optimistic forecasts don’t anticipate large-scale adoption before 2030, and there’s little reason to dispute that timeline.
Let’s be clear: humanoid robots are not ready—at least not yet. For now, they remain more concept than reality, a field brimming with promise but still constrained by technical and economic hurdles. However, this isn’t just another skeptic’s take. Rather than focus on why they aren’t viable today, this report examines what it will take to make them feasible. From AI and mechanics to materials science and cost structures, we explore the critical breakthroughs needed for mass adoption.
China’s humanoid robotics sector is a focal point of this transformation. Now home to more than half of the world’s humanoid robotics firms, China has seen nearly 100 new companies emerge in the past year alone, driven by rising investment and technological advances. Recent headlines might suggest China is catching up—or even surpassing—the West in this field, but the reality is more complex. While its growth is undeniable, China still lags behind in key areas such as AI sophistication, high-precision manufacturing, and software-hardware integration.
That said, the field is still in its infancy, making it difficult to draw clear lines of progress. Humanoid robots, in particular, may not even be the most practical pursuit in robotics. With no clear market leader and no definitive benchmarks for success, it remains uncertain whether China’s surge in humanoid robotics will lead to long-term dominance—especially when it’s still unclear whether humanoid robots are the most viable path for large-scale automation.
Unitree, a key player in China’s robotics push, has drawn significant attention with its viral marketing and ambitious designs. But it represents just one part of a much larger ecosystem. While its approach is eye-catching, long-term success in humanoid robotics will demand more than incremental technical innovation and cost-efficient production. Scalable adoption hinges on robust software-hardware integration and clearly defined use cases that justify R&D investment and the final price tag. That said, Unitree has proven to be a versatile company, and we expect it to pivot as agilely as its world-leading quadrupeds do when commercial opportunities emerge.
In this report we take a deep dive into the competitive race for humanoid robotics in general and Unitree specifically. Are they really as good as the viral videos want to make us believe? The first section of this report, on the progress in humanoid robotics is available to all subscribers, the rest of the report including insights into Unitree is available to paid subscribers. Please consider becoming a paid subscriber yourself (if you are not yet) to support our work and become more informed about China tech.
Enjoy,
Rita Luan, Tech Research Analyst, Co-writer
Rui Ma, Consulting Editor
China’s Embodied Intelligence Ambitions
If you follow Chinese media today, you’ll notice a surge of coverage on humanoid robots and embodied intelligence. While the sector itself isn’t new, its strategic importance has grown significantly. China has now formally placed embodied intelligence on its national development roadmap. The 2025 Government Work Report identified it as a key future industry, alongside biomanufacturing, quantum technology, 6G and more. This wasn’t merely a symbolic mention—it signaled high-level state support and a belief that humanoid robotics could become a transformative economic force in the coming decade.
Embodied intelligence refers to a form of artificial intelligence (AI) in which physical systems—such as robots—interact with their environment to develop and apply intelligent behaviors. Unlike traditional AI, which processes information in digital or virtual contexts, embodied intelligence integrates sensory perception, decision-making, and motor control, allowing robots to operate in the physical world.
At the heart of embodied intelligence are humanoid robots, which combine AI-driven cognition with a physical form, enabling them to perceive, interact with, and adapt to their surroundings—a leap beyond traditional, software-based AI. If this push succeeds, humanoid robots could become as ubiquitous as cars or smartphones, unlocking what many predict will be a trillion-dollar market.
However, achieving mass adoption will require overcoming significant technical and economic hurdles. Unlike traditional industrial automation, where China already holds a dominant position, embodied intelligence presents a far greater challenge. It’s not just about building robots to perform repetitive tasks—it’s about developing machines that sense, learn, and navigate the complexities of real-world environments. This means tackling simultaneous breakthroughs in AI, hardware, supply chains, and large-scale deployment—all at once.
As expected, China is not leaving humanoid robotics to market forces alone. Several major cities and provinces—including Beijing, Shenzhen, Guangdong, Zhejiang, and Shandong—have rolled out policies aimed at accelerating the industry’s growth. Among them, Beijing and Shenzhen have set the most ambitious targets for 2025-2027.
Shenzhen’s strategy focuses on building an industry ecosystem of over 1,200 companies while pushing the sector’s total value past ¥100 billion ($14 billion).
Beijing, meanwhile, is prioritizing mass production, setting a target of 10,000 humanoid robots and at least 100 large-scale applications deployed across industries.
Rather than just developing humanoid robots, China is laying the foundation for an entire embodied intelligence ecosystem. The strategy closely mirrors its approach to electric vehicles—investing heavily in core technologies, securing supply chain dominance, and driving costs down through large-scale manufacturing.
This ecosystem begins with core AI technologies such as machine learning, sensor systems, and decision-making models. It extends into hardware development, including advanced actuators, dexterous robotic hands, and high-efficiency batteries. Software integration is another crucial piece, with companies working on AI-driven control algorithms, real-time computing, and cloud-based intelligence to enhance adaptability.
However, the real test lies in commercial applications. China is targeting a broad range of industries—including manufacturing, healthcare, service sectors, and public safety—where humanoid robots could eventually handle tasks such as factory logistics, rehabilitation care, retail service roles, and security patrols. But as of now, none of these applications are commercially viable at scale.
2024: The Year Humanoid Robotics Took a Leap Forward
In China’s tech industry, it is common to refer to a "yuan nian" (元年)—the first year of a major technological or business model breakthrough. For humanoid robots, 2024 marked that turning point. It was the year when fundamental shifts in AI, control systems, and design philosophy dramatically lowered development barriers, setting the stage for rapid innovation.
Before 2024, most humanoid robots relied on Model Predictive Control (MPC), an optimization-based method used for motion planning in autonomous systems. MPC predicted a robot’s movements based on a dynamic model, optimized control inputs such as joint angles and torque, and executed the first step of the plan before recalculating for the next.
MPC had clear advantages: it was precise, adaptable, and ensured balance and stability, making it a reliable approach for bipedal robots. However, it also had significant drawbacks. It was computationally expensive, requiring real-time optimization at every step. It was also difficult to fine-tune, forcing engineers to manually adjust parameters. Most critically, it resulted in long development cycles, often taking months or even years to produce a well-functioning humanoid robot.
By mid-2024, the industry had moved away from MPC, with Reinforcement Learning (RL) emerging as the dominant approach for robot motion control. The shift was profound. RL allowed robots to learn from experience rather than rely on rigid optimization models, fundamentally changing how they were developed.
Faster Development – Even engineers without prior experience in humanoid robotics could train a functional walking robot in just 60–90 days, an unprecedented improvement.
Lower Barriers to Entry – Companies that previously lacked deep expertise in robotics could now enter the market, significantly accelerating industry-wide innovation.
Improved Adaptability – Robots using RL could refine their movements autonomously through real-world interactions, reducing reliance on pre-programmed motions.
What once took years of iteration could now be achieved in months. The impact of this shift was not just technological—it unlocked new commercial opportunities, drawing a wave of new entrants into the humanoid robotics space.
Another major industry shift in late 2024 was the growing pivot from bipedal humanoid robots to wheeled designs. While legged robots had long been the industry’s ideal, practical considerations led many companies to reassess their approach.
Wheeled humanoids offered several advantages:
Higher Stability – Unlike bipedal robots, which require continuous balance adjustments, wheeled designs avoid complex locomotion challenges.
Energy Efficiency – A bipedal robot consumes power even when standing still, while wheeled robots only use energy when moving.
Better Payload Handling – Wheels allow robots to carry and manipulate heavier objects more easily, making them more suited for industrial applications.
Easier Workplace Integration – Wheeled robots can already perform 80–90% of workplace tasks, without the engineering complexity of human-like walking.
In 2025, wheeled humanoid robots are expected to scale more rapidly in manufacturing and logistics. An unexpected takeaway in a discussion about humanoid robots? Perhaps. But as we’ll highlight throughout this piece, proving real utility is no simple task. A viral video—like Unitree’s striking display of a robot performing kung fu—may capture attention, but true value lies elsewhere. In fact, Unitree’s real strength isn’t even in humanoid robots but in its broader portfolio of robotics, many of which look nothing like a human yet offer far greater practical impact.
GAC Group's self-developed third-generation embodied intelligent humanoid robot, GoMate. Source: Guangzhou Daily
Robotics 101
Before we dive in, let’s clarify a few key points about robots. If you’re already familiar, feel free to skip ahead. But if not, this section will provide essential context for understanding just how far manufacturers still have to go to achieve true human-like capabilities—both in cognition and, perhaps even more challengingly, in physical movement. Spoiler: they’re nowhere close.
Degrees of Freedom Are A Key Indicator of Function
Degrees of Freedom (DOF) define how independently a robot can move, shaping its flexibility and range of motion. In three-dimensional space, a rigid body has six DOF—three for linear movement along the X, Y, and Z axes and three for rotation (roll, pitch, and yaw).
For humanoid robots, DOF corresponds to the number of independently moving joints. More DOF allows for smoother, more human-like movement and greater task complexity. Each joint contributes at least one DOF, with fully articulated models incorporating intricate structures for enhanced dexterity.
Robotic hands are crucial for precise interactions, but developing dexterity remains a major hurdle. The most widely used transmission method is connecting rod transmission, valued for its high precision. However, it requires powerful motors, which add significant bulk to the design.
Current prototypes typically feature around 12 DOF per hand, with future designs expected to achieve greater flexibility as materials and mechanics advance.
DOF Breakdown in Humanoid Robots
Arms – Around 7 DOF per arm, with movement across the shoulder, elbow, and wrist.
Hands – Often 15 to 20 DOF per hand for fine motor skills like gripping and manipulation.
Legs – Typically 6 DOF per leg, essential for stable bipedal walking.
Head & Neck – 3 to 4 DOF for expressive movement and visual tracking.
Torso – 2 to 3 DOF for bending and twisting.
DOF and Functionality
Basic mobility (20–30 DOF) – Sufficient for stable walking and fundamental arm movements.
Moderate dexterity (30–40 DOF) – Greater flexibility in the torso, head, and hands for refined motion.
Advanced dexterity (50–60+ DOF) – Required for delicate operations like tool use and precise manipulation.
Benchmark Humanoid Robots
Boston Dynamics’ Atlas (~28 DOF) – Prioritizes agility and mobility.
Tesla’s Optimus (~28 DOF in early models, increasing over time) – Designed for factory tasks.
Honda’s ASIMO (~57 DOF, discontinued in 2018) – Among the most advanced in gesture and movement.
Shadow Dexterous Hand (~24 DOF per hand) – Near-human hand dexterity, mainly for research.
For truly human-like capabilities—such as assembling electronics, playing instruments, or preparing food—a humanoid robot needs at least 50 to 60 DOF, with continued advancements focused on refining hand dexterity and upper-body flexibility. The kung-fu fighting Unitree G1 that’s been taking over social media? 23-43 DOF depending on configuration. Its H1, the one that was shown in the Spring Gala in a traditional Chinese dance? It doesn’t even have hands (and looks to be <30 DOF.)
That being said, there are some companies who are making more advanced robots with more DOF, see below for some of the more exciting names.
Motion Control Systems Are Complex Yet Primitive
Humanoid robots rely on motion control systems that function like the human cerebellum, executing algorithms for precise, coordinated movement. These systems use a bus connection with separate lines for power and joint transmission, ensuring synchronization even in robots with 20, 40, or more joints. A centralized control system prevents conflicts, allowing for smooth multi-joint coordination.
Motion control systems are a major cost driver in humanoid robots, typically priced between ¥8,000 to ¥12,000 ($1,100 to $1,650). Hardware costs account for about 45%, with CPUs as the most expensive component, while the remainder covers software, including platform purchases, secondary development, and licensing fees. In an ¥8,000 ($1,100) motion adjustment system, for instance, ¥3,000 ($400) is allocated to hardware, while ¥5,000 ($700) covers software.
The CPU works alongside machine learning, image processing, and mapping systems for autonomous navigation and path planning, while additional processors handle vision data, LiDAR inputs, and real-time motion coordination. Due to the high computational demands of motion control—particularly in multi-axis coordination and complex algorithm execution—x86 architectures are preferred over ARM-based CPUs. Many systems integrate GPU-equipped CPUs to handle large datasets and execute complex algorithms efficiently. Typically, three CPUs operate in parallel to manage motion control, environmental perception, and decision-making, forming the robot’s core control center.
Currently, NVIDIA GPUs dominate motion control and visual processing in robotics due to their superior computational power. However, export bans on NVIDIA chips to China are creating challenges for domestic robotics firms. Without NVIDIA GPUs, real-time visual processing and large-scale data training may suffer, forcing companies to seek alternative suppliers or develop in-house chips. Industry experts agree that existing alternatives have yet to match NVIDIA’s performance, potentially leading to higher costs and technical bottlenecks—especially in integrating vision systems with motion control for hand-eye coordination, which demands real-time communication between processors.
However, even with advanced GPU-based algorithms, training a robot to perform a simple grasping action can take 10 to 20 hours before reinforcement learning optimizes the process. Despite their seamless appearance, many public robot demonstrations are pre-scripted to ensure smooth execution. A famous example is Xiaomi’s CyberOne robot handing founder Lei Jun a flower – change it to any other object and it likely would have required dozens more hours of learning.
A positive note is that collaborations between motion controller manufacturers and tech firms are accelerating robotics innovation by improving functionality and shortening development cycles. A key example is Unitree’s partnership with Orbbec, which integrates stereo vision technology into Unitree’s robots, significantly enhancing object recognition and grasping efficiency. Orbbec’s pre-configured data within controllers has also cut development time.
Data Acquisition and Learning
Data collection and learning in humanoid robotics is far more challenging than it seems, even with reinforcement learning on the horizon.
Robots gather data through automated execution and remote control intervention. In automated execution, they follow programmed commands while collecting data to refine their movements. They gather three main types of data: sensor inputs like camera images and environmental scans, motion execution records tracking movement and force, and decision-making data, which helps optimize real-time actions.
To grasp the level of precision in this data, consider how robots track their environment and movements. Basic sensor data, gathered through vision, LiDAR, and IMU readings, provides situational awareness. Motion execution records, captured 100 to 1,000 times per second, track every movement and applied force with extreme accuracy. At an even finer level, motor data, recorded 500 to 1,000 times per second, monitors force distribution and mechanical shifts in real time, ensuring smooth and controlled motion.
When automation falls short, human operators take over, guiding the robot and generating additional data to improve future autonomy. This hybrid approach allows robots to learn iteratively, enhancing their performance over time. And especially in hazardous environments like chemical plants, deep-sea exploration, or space missions, remote operation remains crucial where automation alone isn’t enough. Human oversight ensures safety and precision, allowing robots to operate in conditions too complex for full autonomy.