In recent years, interest in humanoid robots has surged, primarily fueled by the burgeoning potential in this field. Huang Renxun repeatedly stated last year that robots are on the cusp of their "ChatGPT moment," while NVIDIA views robots as the next significant breakthrough beyond generative AI. Conferences like ROSCon for robot developers have gained immense popularity, with Unitree Technology's robots dancing yangko at the Spring Festival Gala and Tesla's Optimus folding clothes, capturing significant attention. However, a humanoid robot marathon held in early April dampened some of the enthusiasm.

Recently, the Beijing Yizhuang Half Marathon and Humanoid Robot Half Marathon commenced at the South Gate of Nanhaizi Park. This was the world's first half marathon for humanoid robots, with 20 teams and running enthusiasts embarking on a 21.0975-kilometer journey to challenge their limits together.

Several online videos revealed that the much-anticipated Unitree Technology G1 humanoid robot fell during the race but subsequently stood up and continued. In response, Unitree Technology issued a statement: "Unitree has not participated in any recent competitions and has been focused on preparing for live broadcasts of humanoid robot fights. The G1 has been sold to numerous customers worldwide since its shipment last year and has been in use for an extended period. Independent teams utilized our robots in this marathon. For example, customers participated with their own algorithms, leading to the presence of multiple Unitree robots at the scene. With independent teams operating them, the robot's performance varies greatly depending on the operator or developer. Like other electronic products, the robot's performance is closely tied to the user. We hope everyone will collaborate to advance the robot industry."

A representative from Unitree Technology further clarified that the G1 robots present at the marathon had their algorithms and hardware modified independently by customer teams. For instance, the City-to-City Technology Team equipped the G1 with a road perception system and enhanced its endurance. Their performance reflects the technical adaptability of the customer team rather than defects in the product. "Similar to other electronic products, the robot's performance heavily depends on the user's debugging and optimization. We hope the industry receives more understanding and tolerance from the outside world."

In what seems to be a response to the questions raised, on April 24, Wang Xingxing, founder and CEO of Unitree Technology, met with the Chief Executive of the Hong Kong Special Administrative Region and presented a comprehensive demonstration of the company's independently developed intelligent robot product matrix. During the technical exchange, innovative products such as quadruped robots and bionic robotic arms showcased advanced functions like navigating complex terrain and grasping/moving objects.

Nonetheless, the current development of humanoid robots significantly lags behind the public's market expectations.

Humanoid robots embody one of the best forms of embodied intelligence, emphasizing agent-environment interaction and autonomous behavior. Embodied intelligence is an intelligent system with environmental perception, intelligent interaction, and planning capabilities, comprising three core modules: perception, interaction, and motion control. Through environmental interaction, autonomous decision-making, and execution planning, humanoid robots achieve task-level interaction.

Humanoid robots' task execution mode mirrors that of autonomous vehicles, both relying on the modules of perception, decision-making, and motion control. They receive instructions through voice recognition and environmental perception, conduct high-level planning like task breakdown, optimize path planning through the cerebellum, and issue decision instructions to drive the servo system for body movement to complete tasks. The large model's general understanding and multi-level reasoning abilities endow the humanoid robot with the core of embodied intelligence, enabling it to better comprehend generalized tasks.

However, from a public competition perspective, current humanoid robots face numerous challenges: they struggle with autonomous direction discrimination, indicating an inability to automatically seek targets based on loaded maps; they have poor endurance, requiring battery or even personnel changes after running a few kilometers, sometimes needing alcohol spraying for cooling; their balance is inadequate, necessitating staff assistance during competitions to avoid falls, and they often cannot stand up by themselves after falling; the robot's head may fall off while its feet continue moving, making the head seem dispensable; their arm-swinging movements are clear but become disoriented after falling; in actual movement, most humanoid robots walk rather than run, lacking the simultaneous lifting of both feet.

Some netizens humorously remarked that current humanoid robots have "underdeveloped cerebellums and completely undeveloped brains." This underscores that humanoid robots are still far from true industrialization and commercialization.

Embodied intelligence was envisioned as one of the paths to AGI (Artificial General Intelligence) as early as 70 years ago. However, due to technological immaturity at the time, it experienced ups and downs. The current wave of embodied intelligence is clearly benefiting from the drive of the third AI wave.

The fundamental difference between this wave of embodied intelligence and past robots lies in intelligence levels. For instance, the concept of robots entering factories to work, hyped by Musk in recent years, was realized decades ago. It involved setting a program for robots to complete prescribed actions in a known environment, exhibiting low and passive intelligence, unable to decide when to act.

In the era of neural networks, true embodied intelligence must be proactive, akin to acquiring world information through perception and converting it into machine-processable forms, such as images into pixel values and text into word vectors. On this basis, the machine makes its own decisions and interacts with the world. This shift implies that robots' market positioning might move from industrial scenarios to daily life, driving the heated market for embodied intelligence.

As the "brain" of humanoid robots, chip performance directly determines their intelligence and capabilities. For example, Unitree employs an industry-leading 8-core high-performance CPU and the NVIDIA Jetson Orin high-compute module, making it more advanced than similar robots. With this support, the G1 can simulate tens of thousands of falls and stands in a simulated environment through deep reinforcement learning (DRL), achieving high-difficulty actions like "automatically standing up after falling" and "single-foot jumping to avoid obstacles."

NVIDIA's high-performance GPU chips, with superior graphics processing, significantly enhance visual information recognition and complex task processing capabilities, achieving a qualitative leap. Domestically, multiple technology enterprises have also made key breakthroughs, successfully transitioning chip technology from the experimental stage to commercial applications.

UBTECH's first-generation Walker primarily uses Intel i7 7500U at 2.7GHz and Intel i5 6200U at 2.3GHz chips, while the upgraded Walker X employs Intel i7-8665U (dual-core, 1.9GHz) + NVIDIA GT1030 graphics card with 384 cores.

Fourier Intelligence's GR-1 is the first humanoid robot with end-to-end environmental perception capabilities, primarily using the I7 13700H CPU.

In 2024, Engine AI's Zhongqing Robotics launched the PM01 humanoid robot, marking the entry of humanoid robots into a new era of heterogeneous computing through the deep integration of the Intel N97 processor and NVIDIA Jetson Orin module. This innovative combination achieves breakthrough improvements in energy efficiency, real-time response, and scalability through CPU and GPU collaborative optimization.

Zhiyuan Robotics is equipped with the Jetson AGX Orin 64GB, boasting extreme AI computing power with 275 TOPS of INT8 computing power, eight times that of the previous-generation Jetson AGX Xavier, supporting the simultaneous operation of multiple complex AI models (such as VLM visual language models and SLAM algorithms).

Recently, Black Sesame Technologies will equip the "Tianwen" robot with a wiser "brain" and "cerebellum" based on the new Huashan A2000 and Wudang C1236 chips. "The A2000 chip aims to equip 'Tianwen' with human common sense and knowledge, achieving human-level decision-making capabilities; the C1236 chip enables parallel processing of AI computing and control tasks, providing reliable support for 'Tianwen's stable and precise operation in complex environments."

The A2000 is defined as a "high-compute chip platform specifically designed for the next generation of AI models," with Lite, Standard, and Pro versions offering performances comparable to 1, 2, and 4 industry flagship chips, respectively. The "industry flagship chips" refer to NVIDIA's OrinX. The A2000 Pro chip is comparable to 4 OrinX chips, competing with NVIDIA's upcoming flagship chip Thor, boasting 1000 TOPS of computing power.

The humanoid robot marathon is just the beginning. Competitions can test the long-term motion and running capabilities of humanoid robots, identifying and resolving issues. However, unlocking more motion forms requires multi-dimensional and comprehensive testing.

From April 24 to 26, 2025, the first Embodied Intelligence Robot Games will be held in Huishan District, Wuxi City. Over a hundred enterprises from across the country will gather with nearly 200 exhibits. After the opening ceremony, the Games will feature competitive events like speed running, cross-country running, football, basketball, dance, and fighting, with intelligence, motion planning, and body design abilities serving as "exam points." Application events will focus on tasks like handling, intelligent grasping, indoor rescue, and innovative scenarios, showcasing robots' comprehensive abilities to complete complex tasks like environmental recognition, grasping, and planning/design from multiple dimensions.

In the speed running event, Unitree Robotics will compete alongside the National and Local Collaborative Embodied Intelligence Robotics Innovation Center (Beijing), the National and Local Collaborative Humanoid Robotics Innovation Center (Shanghai), and robot enterprises from Shenzhen, Shanghai, Xi'an, Chongqing, and other places.

It is understood that besides the speed running competition, Unitree Robotics will also participate in the dance performance segment, alongside robots from Dongguan Benmo Technology Co., Ltd., and Shenzhen UBTECH Robotics Co., Ltd.

Some analysts believe that Unitree Technology's sudden announcement of its participation in the first Embodied Intelligence Robot Games is partly related to its product's performance in the Beijing Robot Half Marathon held in late April. Although Unitree Technology stated that it did not participate in the competition and that the participating robots were purchased by customers for display purposes, without targeted debugging before the competition, the results still generated some public opinion. Participating in the Embodied Intelligence Robot Games may be a manifestation of Unitree Technology reaffirming its technical strength, one of the highlights of the Games.