🤖China's Humanoid Robots, Former Huawei Genius‘ Needle-Threading Robot, and Big Tech Reap AI Rewards

Weekly China AI News from August 19, 2024 to August 25, 2024

Aug 27, 2024

Hi, this is Tony! Welcome to this week’s issue of Recode China AI, a newsletter for China’s trending AI news and papers.

Three things to know

The 2024 World Robot Conference in Beijing kicked off last week with 27 humanoid robots showcased. What do you need to know?
AgiBot, a Chinese AI and robot company founded by a former Huawei’s “Genius Youth” recruit Peng Zhihui, launched five new humanoid robots last week.
China’s tech giants—Alibaba, Tencent, and Baidu—are experiencing an increase in cloud revenue, driven by the growing demand for AI and GPU resources.

China's Humanoid Robot Race Accelerates, but Challenges Remain

What’s New: The 2024 World Robot Conference in Beijing kicked off last week with 27 humanoid robots showcased. This showcase reflects China’s push to lead in next-generation robotics but also reveals the industry’s growing challenges.

By The Numbers:

169 companies are exhibiting over 600 products. 60 new robot models are being released. Humanoid robot displays have tripled compared to last year.
Investment amounts in China’s robot industry increased from RMB 1.98 billion (~$279 million) in 2019 to RMB 9.74 billion (~$1.37 billion) in 2023. The global AI robot market is expected to exceed $19 billion in 2024.
Goldman Sachs projects humanoid robot market could reach $154 billion by 2035.

Show Me Robots: Emerging robot players like UBTECH and Astribot are demonstrating advanced robots with diverse capabilities:

Tiangong 1.2 MAX: Developed by Beijing Embodied Intelligence Robot Innovation Center, this 173cm tall, 60kg robot demonstrated walking and object manipulation at the conference opening.
Walker S Series: UBTECH’s humanoid robot Walker S is capable of tasks like tire inspection and intelligent transport in automotive factories. Walker S is already working in the factories of Chinese EV maker NIO.
Astribot S1: Astribot’s AI robot assistant Astribot S1 showcases
skills in calligraphy, sugar painting, and even martial arts like Wing Chun.
Galbot: GalaxyBot’s humanoid is demonstrated in real work scenarios like item transportation and goods arrangement. The 1-year-old startup has received RMB 700 million from Meituan and SenseTime.
SenseRobot: SenseTime’s AI robot can play chess, Go, and international chess at an advanced level.
TORA Series: Developed by PaXini Technology, Tora features multi-dimensional tactile sensors and dexterous hands.

However, industry experts also caution that significant challenges remain. For example:

Most robots still struggle with complex movements and general-purpose tasks.
High production costs make robots economically unviable for many applications.
Dexterous hands have become a focal point of development, but their necessity is debated.
The big picture is while 2024 is seen as the start of humanoid robot commercialization, widespread adoption may take 3 to 4 years.

Chinese Robot Startup Founded by Former Huawei Prodigy Unveils Suite of Humanoid Robots

What’s New: AgiBot, a Chinese AI and robot company founded by a former Huawei’s “Genius Youth” recruit Peng Zhihui, launched five new humanoid robots last week, including the flagship Yuanzheng A2 series and the open-source Lingxi X1.

Why It Matters: AgiBot’s approach of offering both commercial and open-source robots could accelerate humanoid robot development and challenge rivals like Tesla and Figure.

How It Works: The flagship Yuanzheng A2 series is a 175 cm, 55 kg bipedal robot equipped with AI-powered vision and audio processing, capable of performing intricate tasks such as threading a needle (demo shown above). The series includes variants like the wheeled A2-W and the heavy-duty A2-Max models. Key features of the A2 series include:

Upgraded joint modules with a rated torque of 270Nm;
19 degrees of freedom in the dexterous hand;
Integration of RGBD cameras, LiDAR, and panoramic cameras;
A custom communication framework (AimRT) and software platform (AIMA).

The Lingxi X1 is a compact, open-source robot standing 1.3 meters tall and weighing less than 33 kg.

Constructed with two types of self-developed PowerFlow joints;
Features a modular design for easy customization;
Can utilize a smartphone as its "brain" for cost efficiency.

One More Thing: AgiBot has outlined a five-stage roadmap (G1-G5) for embodied AI development, similar to autonomous driving levels. The company claims to be at the G3 stage, focusing on end-to-end, data-driven training.

AgiBot aims to ship 300 robots by the end of 2024, starting from October. The company is backed by notable investors including HongShan, Hillhouse Investment, BYD, and Shanghai Lingang Economic Development (Group).

China’s Tech Giants Reap AI Rewards, Boost Investment

What’s New: China’s tech giants—Alibaba, Tencent, and Baidu—are experiencing an increase in cloud revenue, driven by the growing demand for AI and GPU resources.

Alibaba is seeing AI-related product revenues more than double year-over-year, with their cloud business reporting a 155% increase in adjusted EBITA to RMB 2.3 billion (~$320 million). This growth is driven by customers renting GPUs for their own AI needs, as well as the broader adoption of AI technologies such as LLMs and machine learning-based recommendation engines.
Baidu’s LLM ERNIE now handles over 600 million daily API calls and generates more than one trillion tokens daily—a threefold increase from just three months ago. This reflects the rising demand for AI-powered solutions across various industries. Its cloud business revenue reached RMB 5.1 billion (~$720 million) in Q2, marking a 14% year-over-year growth. Baidu also amplifies its search engine with advanced AI, similar to Google’s AI Overview. Now 18% of search result pages contain generated content, up from 11% in mid-May.
While Tencent did not break down its cloud revenues, Tencent CSO James Mitchell stated that the demand for renting GPUs from customers has been growing rapidly.

More Spending: These companies have also increased their capital expenditures (capex), with a combined spending of RMB 50 billion ($7 billion) in the first half of the year, nearly double the amount spent during the same period last year, Financial Times reported. This investment is focused on acquiring the necessary infrastructure to support AI development, including processors and high-performance computing resources.

Both Tencent’s and Alibaba’s capex in the first half of the year reached RMB 23 billion (~$3.2 billion), marking a 176% and 123% year-on-year increase respectively.
Baidu has taken a more cautious approach, with capex spending at RMB 4.2 billion (about $580 million) in the first half, up just 4% from a year earlier.
In addition, ByteDance has emerged as the largest Chinese purchaser of AI technology, according to Financial Times.

Why It Matters: Despite US sanctions that limit access to the most advanced AI processors, these companies are finding ways to maintain and even accelerate their AI development. By investing in slightly less powerful but still effective processors, they are ensuring that they can continue to support the training and deployment of AI models. The increase in cloud revenue driven by AI is a clear indication that these investments are paying off.

Weekly News Roundup

Kunlun Tech, primarily known as an online game publisher, has unveiled an AI platform called Skyreels, touted as the “world’s first AI reels platform.” This tool stands out for its ability to autonomously generate scripts, characters, plots, dialogue, and music to produce short videos, each up to 180 seconds in length. (South China Morning Post)
State-linked Chinese entities are circumventing U.S. export restrictions on high-end chips and AI models by accessing cloud services from Amazon Web Services (AWS) and other providers through intermediaries. (Reuters)
Baidu has taken steps to prevent Google’s Googlebot and Microsoft’s Bingbot from indexing content from its Wikipedia-style service, Baidu Baike. (South China Morning Post)
Huawei’s smart car technology unit, Yinwang Smart Technology, has been valued at $16 billion following a 10% stake purchase by Avatr Technology, a strategic move as Huawei diversifies into China's electric vehicle market. (Bloomberg)

Trending Research

A wireless controlled robotic insect with ultrafast untethered running speeds

Researchers from Beihang University developed a novel robotic insect that can achieve unprecedented running speeds without the need for physical tethers, thanks to advanced wireless control systems. This robotic insect showcases a significant leap in the field of micro-robotics, combining agility, speed, and precise control. The design leverages bio-inspired principles and state-of-the-art materials to mimic the locomotion of real insects, enabling a wide range of potential applications from environmental monitoring to search and rescue operations.

Controllable Text Generation for Large Language Models: A Survey

Researchers from Renmin University, China Telecom Research Institute, and the Institute for Advanced Algorithms Research present an in-depth examination of the current state-of-the-art methods and techniques for controlling the output of LLMs during text generation. It categorizes these methods based on various attributes such as the type of control, the stage at which control is applied, and the underlying mechanisms of the models. The survey aims to provide insights into the challenges and trade-offs involved in achieving desired outputs from language models while maintaining fluency and coherence in the generated text.

MeshFormer: High-Quality Mesh Generation with 3D-Guided Reconstruction Model

Researchers from UC San Diego, Hillbot Inc., Zhejiang University, and UCLA present MeshFormer, a novel approach that integrates 3D information to guide the reconstruction process. This method stands out for its ability to generate meshes with superior quality, which is crucial for various applications in computer graphics, vision, and virtual reality. By incorporating 3D guidance, MeshFormer overcomes the limitations of traditional 2D image-based techniques, enabling more accurate and detailed mesh representations.

Recode China AI