The Best Paper team formed a team to defeat the sensory control integrated brain model

Karting, a thing that even humans may not be good at driving, has been conquered by a bipedal humanoid robot??

Starting from “walking to the side of the car” and “getting on the car”, the robot looks at the road, directs itself, and steps on the accelerator.

Control the force with the right foot, continuously adjust the direction with both hands, visually judge changes in the track, and constantly adjust the center of gravity to maintain overall balance.

Wow, multiple points of contact balance, hand eye, foot, brain, and whole body coordination, extremely precise force control, one robot is fully equipped.

The company that came up with this plan, Symbiotic Knowledge and Action, has only been established for over a month now.

And the most astonishing thing is that the bipedal humanoid robot in the video uses an original end-to-end sensing and control integration solution that is leading the industry, breaking through the current mainstream embodied intelligence’s fragmented contradiction of “brain+cerebellum”. It no longer separates perception, decision-making, and operation control into several relatively independent modules, but instead allows a model to output from seeing the world to the robot’s final action.

You should know that end-to-end and Whole Body Intelligence are the most highly regarded technological paths in the academic field of embodied intelligence in 2026. Few of the world’s top laboratories/companies have produced results, yet they have been so skillfully developed by a young Chinese company established less than two months ago?

Who is Symbiotic Knowledge and Action?

01. Tailored Intelligence Best Paper “Tiantuan”
Symbiotic Knowledge and Action was established at the end of June this year, less than 60 days ago.

If the first ‘outrageous’ aspect of this company is that it has been established for just over a month and has directly produced industry-leading sensory control integrated brain model results. The second ‘outrageous’ aspect is the founding team of this company.

The founding team of Symbiotic Knowledge and Action is almost a embodied intelligent Best Paper youth team.

CEO Ding Pengxiang, born in 1995, holds a PhD from Zhejiang University&Xihu University and is supervised by Wang Donglin. Deeply cultivating embodied intelligence, there are a series of dazzling achievements in my resume that can be called “refreshing articles in the scientific research circle”:

The world’s first quadruped VLA (QUAR-VLA, ECCV 2024) and the first humanoid VLA (Humanoid VLA, 2025.02), the core author of ReconVLA (AAAI 2026 Best Paper), the benchmark for lightweight VLA Adapter with 50 million parameters (GitHub Star exceeded 1k within a week, currently 2k+, the highest open source VLA project in China, and the fifth VLA project with over a thousand stars in the world), the founder of OpenHelix Robotics, an open-source community for embodied intelligence (accumulated 5k+Star), and one of the top five cited embodied intelligence PhDs in China.

The other ‘little geniuses’ in the core team are no less impressive.

A team with a Ph.D. per capita, a post-00s generation per capita, and a top-notch skill per capita, which is rare in China (and even worldwide) and has base models on quadrupeds, robotic arms, and humanoid robots. Their achievements include:

Defeating Pi0.5 and winning the CoRL work of the world’s Chinese team’s first Robotics Summit Best Paper;

•IJCNN Best Paper Candidate；

• Proposed the first cross body Whole Body Manipulation Model HEX for full-size bipedal humanoid robots;

The first open-source pedestrian remote control system OpenWBC in the world;

The first Loco manipulation VLA Trajbooster that has attracted high attention in the industry directly “breaks through” the efficiency of cross ontology data utilization

Loco-manipulation、 Four legged, robotic arm, humanoid, WBC counting, pre training, post training, operation control, cross body

Its solid capabilities and leading research have simply put a group of doctoral thesis authors at the forefront of embodied intelligence in China over the past two years into the same company.

More importantly, it’s not as simple as each of them having “published a few papers”. Three out of the top five cutting-edge works in the international Loco manipulation field, as well as a startup team of less than 20 people, have a high density of Best Papers that are somewhat lacking in ethics.

The youngest batch of “paper masterpieces” in China’s embodied intelligence have begun to form groups.

It is interesting that those who closely follow the embodied intelligence track will find that symbiotic knowledge and action is not an isolated case.

In the past six months, the embodied intelligence track has quietly undergone a talent paradigm shift. In the past two years, the most typical entrepreneurial paradigm of embodied intelligence in China was the fate of professors. University laboratories first accumulate technology for a few years, with professors becoming the technical banner of the company, and then enter the industry with doctoral, master’s, and laboratory achievements. Behind a group of celebrity embodied intelligent startups, such a path can be found.

But starting from 2026, more and more young PhDs who are close to the forefront of scientific research are starting to end up on their own.

Why? No one else, the technology is too fast. In 2023, the industry is still discussing whether robots need a “big model”; In 2024, VLA will become a core concept; In 2025, dual system models such as Figure Helix will quickly become popular; By 2026, the industry will be discussing end-to-end.

This has also led to an interesting “knowledge inversion” in today’s industry: the most advanced embodied intelligence technology cognition may not necessarily be mastered by the industry’s “senior” experts. The first people to truly know the next generation technology paradigm may be frontline doctoral students who personally observe data, adjust models, and drop robots every day.

Which plan is just a beautiful benchmark, which method will collapse as soon as it is implemented, and where will the top technology, talents, and resources flow? We have to ask the doctoral students who have just run the model.

This kind of change has actually been demonstrated once in the big model industry. DeepSeek founder Liang Wenfeng once introduced the portrait of the core team in an interview with 36Kr: “They are all fresh graduates from top universities, interns from PhD 4 and 5 who have not graduated, and some young people who have only graduated a few years ago. ”

This is true both domestically and internationally.

Big models are like this, so is embodied intelligence.

02. End to end, embodied intelligence “finale”
The most cutting-edge team and the most disruptive talent structure must naturally produce the most cutting-edge and disruptive results.

As mentioned earlier, the disruptive nature of symbiotic knowledge and action lies in its ability to overturn the current mainstream “layered” approach and directly target the global challenge of end-to-end perception and control integration.

To understand its disruptive nature, one must first know. In the past few years, the mainstream idea of humanoid robots has essentially been “layered”, or “brain+cerebellum”. The upper brain is responsible for perception, understanding, and planning; The lower cerebellum is responsible for motor control and balance.

The Symbiotic Knowledge Action Team is one of the earliest teams in China to propose this architecture, and it released relevant results in 2023.

However, in the past two years, the issue of “layered schemes” has also become increasingly prominent. Due to the natural indirect cascading error between its upper and lower layers, perception and control mapping cannot be directly connected, and the “brain” is extremely difficult to obtain action feedback from the “cerebellum”, resulting in poor task performance.

If we compare human beings: when we pick up a cup that is falling off the table, we will not finish the visual understanding first, then pause to generate the trajectory, and finally call the arm; Often, when the eyes see the cup fall, the body tilts forward and the hand extends out. Perception, movement, and balance are inherently one thing.

For bipedal humanoid robots, this problem will be even more extreme. Every time the robot raises its hand, turns around, or bends over, it changes its center of gravity; Changes in foot contact force can in turn affect upper body movements. Splitting of upper and lower limbs, inability to control lower body and force based on task semantics, long recovery chain from failure, difficulty in unifying data and objectives

That’s also why, in the current demonstrations of most bipedal humanoid robots on the market, it’s rare to see movements that require the cooperation of hands, eyes, brain, and whole body, like the symbiotic “go kart”. It’s not that I don’t want to, it’s that I can’t do it.

The industry has already recognized this issue. Starting from the second half of 2025, “Whole Body Intelligence” and “end-to-end” have become recognized as the next direction in industry conferences, but currently there are very few results that can be achieved.

The goal of symbiotic knowledge and action is in this direction.

The DPC (Direct Perception Control Model) for symbiotic knowledge and action is a Loco manipulation perception and control integrated base model that breaks through the “brain+cerebellum” barrier of robots and directly achieves end-to-end, whole body collaborative operation.

Its input is multimodal information perception data (visual, language, Robot Stat, physical signals); Through the Unified Perception Control Model, full body movements with force feedback are directly generated from visual, linguistic, ontological, historical, and tactile feedback.

Finally, it can directly output the coordinated joint torque of the whole body, aiming at the “finale” of the Interaction Native Whole Body Intelligence track.

Of particular note is that the Symbiotic Knowledge and Action team has taken the lead in verifying that multi-source heterogeneous data training can effectively achieve full body data scaling of humanoid robots. Its ontology free pre training algorithm only requires a very small amount of real data to achieve complex full body collaborative operations.

Compared with traditional solutions, the DPC (Direct Sensing Control Integrated Model) of symbiotic knowledge and action can synchronize and optimize the perception and control signals of both big and small brains (hand eye brain whole body coordination when driving karts), greatly improving the robot’s flexibility control ability (the pressure amplitude of the kart accelerator pedal and the speed of the vehicle body). At the same time, it can achieve universal adaptation of the same base model across ontologies and configurations, and also has strong zero sample generalization ability in real scenarios.

Today, “end-to-end” is the most cutting-edge consensus direction in embodied intelligence research, and it is also recognized by the industry as the evolution direction of the “next generation robot brain” by 2026.

In the past two years, with the maturity of WBC technology in operation and control, the exponential explosion of available data, and the convergence of the VAM perception and decision-making integrated base model route, the “three major mountains” that stood in the way of embodied intelligent perception and control integrated brain have gradually been removed.

In 2026, all three conditions will be met simultaneously. Now is the window period for the end-to-end sensory control integrated brain of bipedal humans, which is about 6-7 months, or even only 4-5 months, and it will not last long, “said Ding Pengxiang, CEO of Symbiosis Zhixing, to 36Kr.

Grasping the window is crucial; Winning once is very important.

But it’s not enough.

03. It’s not just about being able to “win”, it’s about being able to “keep winning”
Every aspect of human society is designed for human configuration, and many scenarios can only be achieved by bipedal humanoid configuration. From a first principles perspective, bipedal humanoid robots are definitely the end of the future, “Ding Pengxiang told 36Kr in an interview.

The goal of symbiotic knowledge and action is not short-term implementation or any easy to pick fruits, but a larger and more universal embodied intelligent AGI future.

This is undoubtedly a very cutting-edge, even very radical future.

But since 2022, the team of Symbiotic Knowledge and Action has almost “bet” on the direction of development of all embodied intelligence technologies, from Humanoid VLA providing task understanding, visual grounding, and action semantic entry; Provide data flywheel for end-to-end Whole body Action Model at TrajBooster; From HEX providing a foundation for Humanoid native State and Action representation for end-to-end models, to gradually integrating Force cerebellar/safety contact experts into end-to-end models through distillation

Making one bet right is a ‘bet’, making ten bets right is not something that can be explained by luck.

Our team has been leading the development of bipedal Loco manipulation, and we are both industry participants and standard setters. ”Ding Pengxiang said.

In 2026, Symbiosis Zhixing is betting on a bipedal humanoid end-to-end sensory control integrated brain model, and tells everyone that this will be the end of the industry.

And their ‘confidence’ lies in continuously cultivating the embodied intelligence industry, never being limited by outdated methodologies, constantly breaking through and exploring; Not just ‘win once’, but also ‘keep winning’.

We can also see the subtle changes in China’s innovation paradigm in them.

For a long time in the past, China’s technology industry was best at engineering, industrialization, and making technology faster, cheaper, and on a larger scale. But in the forefront of basic models and algorithm paradigms, the industry often tends to first look at what has happened in Silicon Valley and then discuss how China can keep up.

In the past few years, this rhythm has started to change.

In the field of big models, DeepSeek has proven that a group of researchers in China with a relatively young average age can also provide their own answers to the world’s most cutting-edge problems and offer breakthrough innovations from 0 to 1.

Now, a similar scene is happening with embodied intelligence, and the protagonists are even younger.

On today’s embodied intelligence track, there are more and more post-00s players. They may still be students in their fourth or fifth year of study, just over 25 years old, but they are already writing the world’s most cutting-edge papers, producing Best Papers, defining the latest model architectures, and driving the development trajectory of the entire industry.

What is truly worth noting behind this is not just “which few new body intelligence startups have emerged in China”, but also the gap between China’s top young scientific research talents and the world’s innovation forefront, which is rapidly narrowing, catching up, surpassing, and even leading the entire industry.

If in the past few years, big models have shown the world for the first time the power of young AI researchers in China, then embodied intelligence may be ushering in the next “world-class moment” for China.

This time, the Best Paper team also collectively went off.