Wang Xingxing has once again spoken very frankly.
Before the opening of the World Internet Conference Wuzhen Summit on November 3, the founder of Unitree Robotics sent a clear signal to the outside world: Unitree's future focus is to conquer robot brain technology, and it does not intend to develop this in isolation. Instead, it seeks global cooperation.
Over the past decade, the company's impression on the public has been almost entirely built on its "body"—extreme hardware cost control, robust motion control capabilities, and driving the price of bipedal humanoid robots to levels far below competitors. Wang Xingxing's own academic background has always centered on mechanics, electromechanics, and automatic control, never straying from the main line of hardware engineering.
Yet this company, which rose through hardware prowess, is now placing the "brain" at the forefront. This shift deserves clarification.
1. Why "Brain Supplementation" Is Urgent Now
Wang Xingxing's urgency regarding the brain did not arise suddenly. In October of this year, at the 2026 World Physical AI Conference, he had already outlined Unitree's strategy in considerable detail: first, scaling production capacity to meet market demand in the range of tens of millions to hundreds of millions; second, advancing the implementation of physical AI small models tailored to their hardware through a parallel approach of self-research and collaboration; and third, deploying ultra-large-scale computing clusters.
By the time of the Wuzhen Summit, he singled out "robot brains" as a key priority, indicating that this task has shifted from being "important but not urgent" to "requiring immediate intensive effort."
The reasons are quite clear. Unitree has already established significant product advantages in "cerebellum" capabilities such as motion control, gait balance, and dynamic movements: The H1's knee joint torque and walking speed match or even exceed those of Boston Dynamics' older Atlas model in terms of parameters, while the G1 has lowered the threshold for bipedal humanoids to the CNY 100,000 level. Humanoid robot shipments exceeded 5,500 units in 2025, ranking first globally; shipments reached approximately 5,900 units in the first half of 2026, with an annual target of10,000 to 20,000 units. Unitree has successfully achieved scalability and cost control at the hardware level.
However, shipment volume does not equate to intelligence. If robots can only perform limited tasks under preset rules and fixed-point motion control, the concept of "general-purpose labor" becomes untenable. Recently, Wang Xingxing judged that the window for "breaking through" physical AI is "one to two years at best, two to three years at worst." This means the time left for Unitree to supplement its brain capabilities is not ample.
2. Tangible Progress Made in Models
If one looks only at the release pace of open-source models, Unitree's progress on the brain front is faster than outsiders might think.
In September 2025, they open-sourced the world model UnifoLM-WMA-0; in January 2026, they open-sourced the Vision-Language-Action (VLA) model UnifoLM-VLA-0; and in September 2026, they released and fully open-sourced the 6B-parameter UnifoLM-WLA-1.0, making code, models, and datasets simultaneously available.
The key significance of WLA-1.0 lies in addressing a practical issue with the VLA route. Traditional VLAs function more like "seeing and acting": observing objects, understanding instructions, generating actions, and reacting to the current frame. However, the real world is dynamic, where one action changes the environment for the next moment. Unitree's approach is to separately learn "where changes will occur," first predicting changes in interaction zones, then generating actions, thereby integrating embodied reasoning, future change prediction, and action learning into a unified system. Simply put, it determines what will happen before deciding how to act.
From the results, this route has at least secured entry qualifications in benchmarks. UnifoLM-WLA-1.0 achieved 7 open-source State-of-the-Art (SOTA) scores in embodied reasoning evaluations, with scores in certain spatial understanding tasks—such as Where2Place and EmbSpatial—exceeding those of GPT-6 Astra. On actual robots, a single model coordinates 64 tasks, covering 54 desktop operations and 10 whole-body operations, ranging from folding towels and organizing tableware to making beds, taking out trash, and loading clothes into washing machines, supporting both two-finger grippers and various five-finger dexterous hands.
These tasks are still far from being "general-purpose," but the direction is clear: enabling one model to cover more tasks and more end-effectors, rather than customizing solutions for each scenario individually.
3. Funding and Talent Are Moving, But Team Background Remains a Practical Challenge
Developing the brain is capital-intensive. Unitree's R&D investment for the first three quarters of 2026 was CNY 263 million, a year-on-year increase of 191.48%. Specifically, Q3 R&D investment was CNY 127 million, up 249.58% year-on-year, with the increment primarily driven by embodied large model development, computing power investments, and team expansion. Of the CNY 4.2 billion raised from its IPO, half is planned for "robot brain" R&D, covering iterations of embodied intelligence large models, dexterous hands, motion control optimization, and autonomous operating systems.
Funding is one aspect; the talent issue is more pressing. Unitree's core technical team has long been dominated by backgrounds in mechanics and motion control, with AI algorithms, particularly deep learning and large models, added later. Starting in 2025, Unitree significantly expanded recruitment for "embodied large model" related roles, with data infrastructure positions marked as "hotly recruiting."
At the IPO celebration ceremony, Wang Xingxing mentioned the "brain" pre-research direction for the first time: "Physical AI Self-Evolution." This relies on large models to enable robots to achieve an autonomous loop of perception-decision-execution-evaluation-learning-evolution. If this closed loop can be operationalized, its significance is profound—robots would no longer rely on engineers for item-by-item tuning but could improve themselves through real-world interactions.
However, "self-evolution" remains a pre-research concept, distant from engineering verification. Unitree's actual status on the brain front is that the direction is set, the path is laid out, but core capabilities are still developing.
4. Cooperation Is Not an Alternative, But a Strategy
At Wuzhen, Wang Xingxing placed "seeking global cooperation" and "conquering robot brains" in the same sentence. Two aspects of Unitree's brain-related cooperation layout are already quite clear.
The first is the "body-brain" division of labor with NVIDIA. In June 2026, NVIDIA announced a partnership with Unitree to launch the humanoid robot reference design H2+, based on the Isaac GR00T system: NVIDIA provides the Jetson Thor computing platform and Cosmos model foundation, while Unitree supplies the engineering-validated H2 body, jointly constituting a "brain" that drives full-body motion, multimodal perception, and autonomous decision-making. In the short term, Unitree leverages NVIDIA's computing power and foundational model base, avoiding building all upstream capabilities from scratch. However, the other side is also clear: if they remain long-term in the division of "NVIDIA builds the brain, Unitree builds the body," the center of value distribution will not lie with Unitree.
The second is the strategic alignment with DeepSeek. In August 2026, during the IPO roadshow, Wang Xingxing revealed that they had signed a Strategic Cooperation Memorandum with DeepSeek, focusing on Artificial General Intelligence (AGI), high-performance robotics, and AI large models. This line offers greater imaginative potential but is also at an earlier stage.
The strategy involves using cooperation to secure the time window while simultaneously pushing self-developed models and open-source ecosystems as long-term leverage. The full open-sourcing of the UnifoLM series carries the implication of "Android for smartphones"—opening the underlying layer allows developers to adapt different devices and scenarios around Unitree's model ecosystem, trading ecosystem influence for voice. Whether this path succeeds depends on two variables: whether the models can continuously iterate to become sufficiently user-friendly, and how many developers are willing to truly treat Unitree's foundation as their starting point.
5. How to Assess the Progress?
Returning to the initial question: How should we judge Unitree's progress on robot brains?
A pragmatic assessment is: The strategic pivot is complete, resource investment is in place, and tangible, verifiable phased results have been achieved at the model level. However, the "brain" as a capability system capable of scaling, generalization, and supporting general tasks remains in the early construction phase.
Unitree has proven that it can run models on actual robots, open-source models that perform well in benchmarks, and use cooperation to gain leverage on computing power and foundational models. What it has not yet proven is whether this "brain" can enable robots to consistently generate reliable economic value in unstructured environments, rather than just completing 64 carefully choreographed tasks in demonstration videos.
In "supplementing the brain," Unitree faces greater challenges. Feedback from motion control is immediate and physical—success is success, failure is failure. Feedback from the brain is statistical and scenario-dependent; good performance on 64 tasks does not guarantee immunity from failure on the 65th task.
Wang Xingxing understands this, hence he defines the "brain" as a priority, brings "global cooperation" to the table, and simultaneously initiates internal self-evolution pre-research. Doing these three things concurrently indicates there is no shortcut for breakthroughs in a single area. The true看点 (key point) of Unitree's brain supplementation lies in whether it can turn the brain into a second barrier, equally difficult to bypass, before its hardware advantages are caught up with by competitors.