Embodied intelligence has become one of the hottest industry tracks today. The listing preparation of China’s industry leader, Yushu Technology, is advancing vigorously, while US-China tech giants are quietly expanding their more ambitious layouts beneath the surface.
At the GTC 2025 keynote speech in March this year, NVIDIA CEO Huang Renxun showcased the humanoid robot NEO Gamma from Norwegian robotics company 1X. This robot uses a post-training strategy built on NVIDIA’s GR00T N1 model and has successfully completed autonomous organization tasks. Through this, Huang Renxun conveyed the core message: the future of humanoid robots depends on adaptability and learning ability — whether a robot can perform tasks hinges on whether its “brain” can adapt to the environment and acquire new knowledge.
This judgment was soon validated. Within a few months, global tech giants delivered new answers in the field of “robotic brain.”
In August, NVIDIA launched the Jetson AGX Thor edge computing platform, which can run multiple generative AI models simultaneously on a robot. It is recognized by the industry as the “new brain for robots.” As a new generation of technology stack for robots and physical devices, this product,Abbreviation “Thor,” is based on the Blackwell GPU architecture, with a peak computing power of 2070 FP4 TFLOPS, a 7.5-fold increase in AI performance compared to the previous Jetson AGX Orin module, and a 3.5-fold improvement in energy efficiency. Its biggest breakthrough lies in enabling robots that previously relied on cloud processing or multi-chip collaboration to achieve real-time perception and decision-making on a single compact module. Currently, early adopters like Amazon’s warehouse robots and Boston Dynamics have begun integrating Thor, aiming to create smarter, more independent robotic products.
The other side of the Earth, ByteDance, has also been very active. In July and September, ByteDance successively launched its self-developed general-purpose robot models GR-3 and Robix, demonstrating the ability to perform complex tasks in real household scenarios. At the end of July, ByteDance’s Seed team released a demonstration video: a robot equipped with the latest robot VLA Achievements GR-3, accurately completed the action of inserting a hanger into a shirt and hanging it. Last week, the Seed team introduced a new achievement, Robix, which, together with GR-3, forms ByteDance’s new generation of robot model matrix.
From Huang Renxun’s GTC 2025 speech to now, it’s been less than half a year. The “resonance” of the two giants in the robotics field makes one wonder: Huang Renxun and Zhang Yiming, this time they thought of the same thing.
A NVIDIA: The Ecosystem Moat of the Hardware King
Before Thor made its debut, NVIDIA was already a leader in the robotics field. Its Jetson platform dominates the development of high-end robots and autonomous robots, with an ecosystem encompassing over 2 million developers and more than 7,000 companies adopting the previous generation product, the Orin series.
In August this year, Thor was officially launched. Judging from hardware performance and maturity, it is almost in an absolute leading position among similar products. Competitors like Intel and Qualcomm, in their edge AI products (i.e., AI running locally on the device), are relatively lagging and still cannot achieve the same level of integrated computing in a single module.
A more critical advantage lies in NVIDIA’s deep integration of hardware and software stack, and leveraging the widespread use of the CUDA toolkit to build a solid ecosystem moat. Thor can directly call NVIDIA’s full Isaac robotics software platform, AI model library, and simulation tools to achieve end-to-end deep integration, including the latest open-source pre-trained foundational model for robotics scenarios — Isaac GR00T N1.
GR00T N1 adopts a dual-system architecture inspired by human cognition: the “fast thinking and acting model” simulates human reactions and intuition, while the “slow thinking model” can reason about the environment and instructions and plan actions. Bernt Børnich, CEO of 1X, stated bluntly at GTC 2025: “While we were developing autonomous models, NVIDIA’s GR00T N1 significantly enhanced the reasoning and skill capabilities of robots. We completed the deployment of NEO Gamma using only minimal post-training data.”
However, Thor is not without flaws. First, the price barrier is high — the Jetson AGX Thor development kit costs $3,499, making it difficult to be a choice for ordinary consumer products. Second, it has higher power requirements, meaning it is more suitable for products with stable power supply capabilities, such as autonomous vehicles, factory robots, and delivery robots.
ByteDance: The “Brain Puzzle” for Chinese Players
On the “Robot Brain” track, NVIDIA quickly welcomed Chinese competitor ByteDance.
At the end of July, ByteDance released the next-generation robot VLA model GR-3. In the official demonstration, the ByteMini robot equipped with GR-3 completed complex tasks such as hanging a shirt on a hanger, picking up household items and placing them in designated locations, and could distinguish item sizes, accurately executing instructions like “pick up the larger plate.”
Analysis of the demonstration and technical reports show that the core advantage of GR-3 lies in its ability to understand complex and abstract language, as well as its strong few-shot adaptation capabilities. According to foreign media reports, the Seed team adopted a hybrid training method: first inputting massive amounts of image and text data into GR-3, then fine-tuning through human-robot interaction in a virtual reality environment, and finally having it learn to imitate robot movements in the real world. This strategy has given GR-3 the ability to adapt to complex and unpredictable environments.
It is worth noting that GR-3 is quite competitive in terms of parameter scale and actual performance. Chris Paxton, an AI scientist at Meta, disclosed in his research on robot VLA that the Byte 4 billion parameter GR-3 model seems to outperform the NVIDIA GR00T with about 2 billion parameters. Reasonably 推测, with sufficient data and computing power, the “scaling law” will still play a role.
However, GR-3 focuses more on the execution level, resembling the “neurons” of a robot’s body, and has not yet fully covered all functions of the “brain.” Meanwhile, the latest achievement of the Seed team, Robix, fills in another piece of the “robot brain” puzzle. The recently released Robix is specifically responsible for core functions in robot systems such as task planning, reasoning, and natural language interaction. Dong Heng, a researcher at Seed, described it on his personal homepage: “This is a unified robot brain that integrates reasoning, planning, and natural interaction, with performance better than GPT-4o and Gemini 2.5 Pro.”
Robix is not a complete “robot brain.” According to the technical documentation, the “body motion / execution” part needs to be completed by a low-level controller model, i.e., VLA models like GR-3. In other words, GR-3 and Robix need to work together to drive the robot. This working logic is similar to the dual-system model of NVIDIA GR00T N1 — both have one system responsible for reasoning and another for execution functions.
In official tests, Robix equipped with GR-3 as the controller completed tasks such as dining table cleaning, cashier packing, and beverage sorting using the ByteMini robot. Compared to other models, except for the beverage sorting project which was slightly inferior to Gemini 2.5 Pro, it achieved the highest score in the remaining projects.
The “Invisible Layout” of Byte: From Technical Accumulation to Ecosystem Investment
Robix and GR-3 are not Byte’s first foray in the field of robot VLM. Although the outside world often views Byte as an internet content company, in the robotics field, it is actually a low-key “invisible giant.”
Byte’s robotics exploration can be traced back to 2020, when Zhang Yiming showed strong interest in robotics and subsequently participated in project discussions periodically. In December 2023, the Seed team launched the GR-1 model as a forward-looking experiment in the field of robot VLM, adopting a “large-scale video generative pre-training + robot data fine-tuning” model. Based on this accumulation, the GR-2 launched in 2024 used a pre-training foundation of 38 million videos and over 50 billion tokens, achieving an average success rate of 97.7% in over 100 control task tests after robot trajectory fine-tuning. And GR-3 and Robix are indeed Byte’s latest technological extensions in the era of large models.
At the technical implementation level, Byte has quietly achieved mass production breakthroughs — currently producing over 1,000 wheeled logistics robots, focusing on an integrated solution for “warehousing + automated transportation,” capable of self-learning and route planning, serving TikTok e-commerce warehouses and external clients such as SF Express and BYD. However, these logistics robots are more the result of early technical accumulation. From the development paths of Robix, GR-3, and ByteMini, it is clear that Byte’s goal is to seize a leading position in the field of embodied intelligence.
Talent deployment is accelerating in sync. Recently, ByteDance’s recruitment website has listed numerous robot-related positions, some of which explicitly mention “next-generation general-purpose robots,” all belonging to the Seed team, with work locations set in Beijing and Shanghai. According to a July report by the South China Morning Post, the Seed team’s workforce is expected to exceed 300 people this year.
In terms of ecosystem investment, ByteDance has also been active. When the domestic robotics and physical intelligence leader, Yushu Technology, completed its C round of funding, the Jinqiu Fund, which has a deep connection with ByteDance, appeared on the list of investors. This fund was founded by Yang Jie, the former head of financial investment at ByteDance, in 2022, with many core members coming from ByteDance’s investment system. The name “Jinqiu” directly originates from Jinqiu Garden in Haidian District, Beijing, where Zhang Yiming and ByteDance started their entrepreneurial journey.
D-Sync After: The Possibility of Giant Collaboration?
ByteDance is accelerating its robot field deployment both internally and externally, but its current technical Accumulate is primarily concentrated on the “robot brain” level on the model side; while NVIDIA’s new-generation chip solution perfectly complements this with its own advantages.
In fact, ByteDance has been a key client of NVIDIA in China for many years, and Huang Renxun is also well aware of the critical role Chinese enterprises play in the field of embodied intelligence. In July of this year, when Huang Renxun attended the opening ceremony of the Beijing Chain 博会, he emphasized: “The next wave of AI will be robotics. Future robots will not only be able to reason and execute but also truly understand the physical world.”
For NVIDIA, the strategic significance of the Chinese market is self-evident. Its official blog shows that several domestic companies, including United Imaging Healthcare, Wanji Technology, UBTECH, and Yushu Technology, have already started using Thor, though ByteDance has not yet appeared on the list. Wang Xingxing, CEO of Yushu Technology, commented: “Jetson Thor has brought a huge leap in computing power, enabling robots with greater agility, faster decision-making, and higher levels of autonomy, which is crucial for robots to navigate and interact in the real world.” At the CES exhibition in January 2025, Huang Renxun appeared on stage with 14 partner humanoid robot companies, six of which were from China, including Yushu Technology and XPeng.
Meanwhile, ByteDance’s Seed team’s layout is not limited to the model domain. ByteMini, launched alongside the testing of GR-3 and Robix, may appear to be an experimental test product, but its technical specifications are impressive —— it includes 22 degrees of freedom and features a spherical wrist design, demonstrating strong operational capabilities in confined spaces and high-dexterity tasks. The introduction of this product already highlights ByteDance’s ambition to develop next-generation embodied intelligence products.
Now, NVIDIA has unveiled a new generation of robot chip solutions, and Huang Renxun and Zhang Yiming have reached a consensus on the track choice. These two tech giants, along with the companies behind them, may have the potential to continue their previous collaboration in the field of robotics.
