While browsing social media at home over the weekend, a video truly impressed me.
A robot driving a go-kart sped along the track, navigating corners, accelerating, and steering—all captured in a single continuous shot.
I didn't expect our robots could already be showing off their skills on a go-kart track!
Wait a minute?
This isn't the same style I saw at the WRC exhibition!
The robots at the exhibition were walking, with sparks and lightning effects. When did the robots in the video secretly sign up for a racing training course behind our backs?
It drove like a complete pro, handling high-speed cornering and agile obstacle avoidance smoothly in one go.
Compared to previous robot demos where they slowly walked over and bent down to pick something up, the robot in this video actually got into the car by itself, achieving hand-eye-foot synchronous coordination and accomplishing multi-contact-point balancing and precise force control at full speed.
Such full-body coordination of a bipedal humanoid robot—where its eyes, hands, and feet work together to perform detailed operations during high-speed motion and complex posture changes—has rarely been seen in previous humanoid robot demonstrations.
The company that released this video is called GongShengZhiXing (Symbiotic Cognition).
What GongShengZhiXing is developing is an end-to-end sensory-motor integrated "brain" for bipedal humanoid robots—the foundational model.
The founder, Ding Pengxiang, was born in 1996. The rest of the core team members are all born after 2000, are all current PhD students, and the company was founded just two months ago.
Regarding this seemingly "unorthodox" demo, Ding Pengxiang explained:
I want robots to be more like humans, not just perform some very mechanical service functions. For example, if a human can drive a go-kart, a robot should be able to do it too.
A question naturally follows: Why would a group of young people who haven't even received their graduation certificates dare to tackle the hardest and most cutting-edge track in embodied intelligence?
Betting on the Hardest Path: Moving Beyond Hierarchical Research to Challenge End-to-End
How new is the track of "end-to-end" bipedal humanoid robot models?
So new that there are no ready-made talents available in the market.
According to Ding Pengxiang, this direction only started gaining traction in the last couple of years. Those who understand the technology are still PhD students, making it almost impossible for them to find people who meet their needs from the existing talent pool.
Those who graduated years ago can't keep up with the pace, and those who haven't reached the PhD level lack a deep enough understanding of the fundamentals. The most cutting-edge breakthroughs are precisely in the hands of this batch of top PhD students.
The team at GongShengZhiXing is exactly such a group of young and cutting-edge individuals.
They have won the most prestigious academic honors in the domestic embodied intelligence field—two Best Paper awards, with one selected as a Best Paper Candidate. They have also built the first domestic embodied foundational model whose GitHub Stars exceeded 2K.
More importantly, they are not followers who jumped on the bandwagon after the trend arrived. They are among the earliest pioneers in China exploring embodied foundational models and are the most steadfast long-termists in this track. The team has collectively published over 40 top-tier conference papers, with their technical footprint covering the full stack of perception, decision-making, control, data, and systems. A series of industry-first works, from dual systems and lightweighting to multi-configuration, all originated from this young team.
Their judgment is clear: The ultimate and most widespread form for robots to enter human life is the bipedal form.
Following the evolution paths of autonomous driving and large models, end-to-end is the most efficient technical paradigm.
This judgment stems from their firsthand research experience.
As early as December 2023, Ding Pengxiang began equipping quadruped robots with "brains."
He released QUAR-VLA as the first author, which was the first VLA (Vision-Language-Action) task paradigm for quadruped robots. It allowed the upper-level model (the brain) to understand vision, language, and task intent, then handed the decisions to the underlying motion system (the cerebellum) for execution.
This layered approach—"the brain manages decisions, the cerebellum manages motion execution"—is what the industry calls the hierarchical scheme, a method still used by companies like Figure today.
During subsequent research, he gradually discovered issues.
He describes the hierarchical approach as a "theoretically feasible but less elegant solution":
The brain and cerebellum work separately. For every new task, the upstream output needs custom fine-tuning to achieve good performance.
Deeper, there's an inherent structural defect.
The brain and cerebellum are trained separately and combined during deployment. Even if each part is optimized locally, it doesn't guarantee global optimality.
Every translation between the two layers adds cascading errors and information loss.
Ding Pengxiang concludes, "As long as it's hierarchical, there is an information bottleneck in the middle. This interface design determines that hierarchical architecture cannot achieve true Scaling."
This conclusion also comes from his long-term observation of the development of autonomous driving and large models.
Early autonomous driving relied on the coordination of multiple modules like perception, prediction, planning, and control. Later, it began exploring models that directly learn driving capabilities from data.
After Tesla's FSD switched to an end-to-end approach, its performance improved significantly.
In Ding Pengxiang's view, robots may undergo a similar transformation, allowing models to learn the complete "perception-understanding-action" process from data.
His logic is, having thoroughly explored the hierarchical path repeatedly since 2023, if it's judged that future data scaling will ultimately push systems toward end-to-end, there's no need to take another detour.
This shift is happening in 2026.
Google's Gemini Robotics 2 already uses a single policy to unify full-body actions from feet to fingertips into one model.
Yet, the consensus at the RSS 2026 top conference clearly states that a single end-to-end model cannot cover the complex full-body dynamics, and hierarchical modularization is the optimal practical solution at the current stage.
The pure end-to-end path that GongShengZhiXing is betting on is a more radical and less-trodden one.
Why bipedal specifically? Ding Pengxiang's answer is first principles: The buildings, tools, and environments of human society are all built according to the human body's configuration.
Wheeled robots can only work fixed in production lines and shopping malls. Bipedal robots can go outside, climb stairs, drive cars, and perform different tasks across scenarios.
Versatility means cost amortization. Bipedal forms also offer an anthropomorphic sense of affinity.
Timing is equally crucial. Before April 2026, there wasn't even a general teleoperation model for bipedal humanoids. Without a data source, you couldn't train a foundational model.
It wasn't until NVIDIA open-sourced Sonic that the industry gained the cornerstone for large-scale data production.
This direction is no longer in the early "want to train but have no data" stage, yet it hasn't matured enough for the technical route to fully converge. GongShengZhiXing is gambling on this exact window in between.
The Technological Watershed: Others Learn Kinematics, They Tackle Dynamics
First, let's clarify a watershed: Most current robot foundational models actually learn kinematics.
Kinematics concerns moving from point A to point B—where the hand moves and whether it grasps. This works fine for fixed-base robotic arms because they don't fall over.
When a bipedal robot reaches out to grab something, its body's center of gravity changes simultaneously.
When squatting to pick up an object, the waist must lean forward, and the ankle joints and legs must redistribute force to maintain balance.
Add friction, collisions, inertia, and contact into the mix, and the model is now facing a full-body dynamics system.
According to Ding Pengxiang, a g1 bipedal humanoid has 29 degrees of freedom, which is far more complex than the 7 degrees of freedom of a robotic arm.
In the past, robots learned "motion trajectories." Humanoid robots need to learn "how the body acts in the physical world."
Moving from kinematics to dynamics, from completing tasks to maintaining stability—this is the real technological watershed for bipedal robot models.
This explains why GongShengZhiXing pushes its model all the way to the Joint Target level.
In hierarchical schemes, the brain first outputs kinematic targets, and the downstream cerebellum then calculates them into joint movements.
GongShengZhiXing removes this intermediate step, making the model directly face the body's state with dozens of joints.
A question arises: With the cerebellum removed, can the robot still stand stable?
Relying solely on imitation learning, the model only learns standard actions. It hasn't seen various stumbling body states during execution, so encountering unfamiliar states might lead directly to a fall.
What it lacks is the ability to recover from failures and stand firm.
The core of GongShengZhiXing's approach is a dual-domain collaborative optimization mechanism called "Task Behavior Modeling — Motion Prior Distillation."
One optimization pathway continues with behavioral cloning, ensuring task accuracy and motion fitting capability. The other, via DriftDistill, transforms the stability, disturbance resistance, and Failure Recovery capabilities accumulated by the underlying controller into the endogenous motion priors of the unified model.
It's not simply adding two Losses. Instead, it integrates the two types of capabilities—"accurately completing tasks" and "stably controlling the body"—into the same model, allowing the large model to simultaneously acquire task intelligence and body intelligence.
If DriftDistill can continue to scale with the model and data volume, it attempts to solve an even bigger problem:
Beyond cognitive Scaling, can a robot also achieve Scaling in its motion control capabilities.
Ding Pengxiang refers to this motion capability aggregation module as the Motion Expert Model, currently sized close to 1B parameters, continuously absorbing capabilities from different motion control models through distillation.
Taking another step forward, the end-to-end large model must also solve the problem of the robot's output force.
Many robot systems today primarily control position. But when a robot truly enters the real world, correct position doesn't guarantee task completion.
A hand reaches the drawer handle but doesn't know how much force to apply—the door still won't open. When handing something to a person, the position is correct, but excessive force is equally unsafe.
GongShengZhiXing's technical route is to first use the Force Expert as a safe contact specialist, letting it learn force application, compliance, impedance, and contact feedback. Then, through policy distillation, it is gradually integrated into the Motion Expert and ultimately into the end-to-end model.
Position determines whether the robot can "get there," while force determines whether it can truly "get the job done well."
This is also the meaning of GongShengZhiXing's "Full-Body Physical World Model": to make the model understand how the body is subjected to forces, loses balance, and makes contact, then directly translate these constraints into actions.
Data, Emergence, and a Calm Timeline
Once model training truly begins, the first major hurdle is the data shortage.
Globally, compliant robot data amounts to only about 500,000 hours as of early 2026, less than one twenty-thousandth of that used for large language models.
The awkward part is that much of this data captures stationary operations—robots working in place within cubicles—offering limited help for full-body mobile manipulation.
GongShengZhiXing's Blog has released around 10,000+ hours of data, with several thousand hours collected by themselves.
What full-body coordination models truly lack is strongly-coupled full-body mobile operation data from real-life scenarios, such as riding a bicycle, driving a go-kart, inflating a basketball, etc.
Their solution is TrajBooster: Extract end-effector trajectories from robotic arms or wheeled robots, have the humanoid robot track this trajectory in simulation while maintaining balance, unify actions from different sources into the same space, reuse the task diversity already defined for robotic arms, and provide low-cost data for pre-training.
When discussing model capabilities, Ding Pengxiang rarely uses the term "generalization."
In his words, "That term isn't very substantive; it's as empty as saying someone is good."
What he wants is "Emergence": At the joint angle level, for instance, if the model is trained on walking and jumping data, it might combine these to produce the new skill of running during testing.
Because only at the joint level can partial motion capabilities potentially combine to form entirely new motions.
Regarding the timeline, his judgment is quite sober.
He infers that around 100,000 hours of data will likely still yield more lab-level demos. Truly meaningful large-scale commercial significance might only emerge after reaching the million-hour mark.
Right now, the industry hasn't even converged on the most suitable data form; routes like EMG, exoskeletons, motion capture, and Pico teleoperation are still competing.
In his view, the three most difficult problems for humanoid robots currently are: first, how perception and control should truly be combined; second, what type of data should be used; and third is the data scale itself.
Ding Pengxiang hopes that one day, a robot model architecture designed by a Chinese team can also be widely adopted by peers in North America.
We don't want to be followers; we want to lead technological development.
Let's return to that go-kart.
What enabled it to navigate corners at full speed was a group of young people who haven't even graduated yet.
They are betting on the hardest and most cutting-edge direction in embodied intelligence.
They are gambling that end-to-end will eventually replace hierarchical approaches.
The industry is far from being able to prove that the final route has been validated. Even Ding Pengxiang himself admits that the entire bipedal foundational model industry has not truly converged yet.
But sometimes, the cutting-edge nature of a startup team lies not in executing a mature answer faster, but in whether it can see the next truly important problem earlier than its peers and dare to start solving it when the answer is still unclear.
This article is from the WeChat public account "QbitAI" (ID: QbitAI), author: Qiao Buchi.
