Yanxi Dynamics Releases Embodied Foundation Model Paimon-0, Tops Four Leaderboards Including RoboColiseum

What a contrast.

A company founded in Shenzhen just three months ago clearly positions itself as building embodied chips, yet the first result it has shown the public is an embodied foundation model (series) called Paimon.

And the first model in this series, Paimon-0, took four first places on its debut:

First, look at the most prestigious one, RoboColiseum. This leaderboard, jointly built by nearly 20 universities, companies, and research institutions worldwide, including Hugging Face, NVIDIA, and Tsinghua University, has recently been highly popular in the embodied AI community; Paimon-0 ranks first with a score of 0.7754.

You should know that this leaderboard recently just invited general multimodal large models such as GPT-6 Astra, Claude Opus 5.5, and Gemini 4 Argon to compete on the same stage:

Although the large models do have impressive scores in instruction following and spatial understanding, when it comes to the general manipulation sub-leaderboard that requires real hands-on action, Paimon-0's 0.6613 is still the highest overall, and it also leads the perturbation adaptation sub-leaderboard.

RoboCasa365 is a large-scale multi-task benchmark for general robot policies, containing 2,500+ simulated kitchen scenarios, 365 daily tasks, and 2,000+ hours of interaction data. It selects 50 tasks and tests generalization ability at three levels: atomic visible, compositional visible, and compositional unseen. Paimon-0 tops it with a composite score of 58.1.

VLA-Arena is an open-source evaluation leaderboard for general embodied models, conducting systematic assessments around safety and anti-interference capability and supporting cross-model comparison. Paimon-0 achieved an average success rate of 0.73, ranking first.

At the same time, Paimon-0 participated in the IROS Workshop challenge hosted by Primotion. The track focused on dual-arm robot policies and included both simulated online evaluation and real-machine offline validation. Yanxi Dynamics ultimately won the championship with a total score of 46.599, verifying the capability of end-to-end policies in dual-arm manipulation.

And the contrast does not stop there.

Before today, this company's results had not been made public (and there was even little news related to the company itself), yet its solution already had customers placing orders!

Specifically, a customer sent two components of a robotic arm to their laboratory. After switching to its control solution, a set of same-machine acceptance records showed that, at a 1kHz control cycle, the worst arrival-time error of the EtherCAT control cycle dropped from 30 microseconds in software mode to 152 nanoseconds in FPGA hardware mode, reduced to about 1/200 of the original.

In software mode, control timing is scheduled by CPU software and is prone to jitter caused by system interrupts and task switching; FPGA hardware mode pushes time-deterministic control down into hardware logic, so the error is greatly narrowed.

Even the customer did not quite believe it at first, but after repeated verification, that is exactly what the facts showed.

So who exactly is this dark horse from Shenzhen? Without keeping you in suspense, it is Primotion.

Although the results are worthy of recognition, questions follow:

Why would a chip company first deliver a model result? And how did it achieve these results after being founded only three months ago?

With these questions, we had an in-depth conversation with the Primotion team.

To answer this question, we first need to look at what kind of solution Primotion plans to build around embodied chips.

Primotion's embodied intelligence software platform connects customer models, open-source models, and the Paimon reference model through open model interfaces, covering the entire process of training adaptation, deployment optimization, and production inference.

On the training side, data processing pipelines and model post-training adapt to specific tasks; the deployment toolchain provides quantization and distillation, compilation and inference optimization, and real-machine validation and evaluation. On the execution side, the cloud-based Meta Brain handles task planning and failure recovery, while the on-device Local Brain performs policy inference, and the embodied runtime coordinates data synchronization and action execution.

Runtime feedback is evaluated by the self-developed Progress Model Paimon-Eval and then enters the RSI self-improvement process, driving continuous model iteration, with real2sim providing simulation assistance for this process. The underlying layer is supported by the embedded software stack and compute-control hardware, connecting to the robot's local control loop.

Primotion wants to connect these links and, around its compute-control integrated SoC and domain controller, form a software and hardware solution that can be delivered to customers.

This approach has precedent in the autonomous driving field: NVIDIA supports upper-layer algorithms landing on chips with vehicle hardware together with system software, middleware, and inference tools, and the DRIVE platform is a typical reference. Migrating to the robotics track likewise requires first clarifying the logic of algorithm execution, data flow, and task collaboration before it can guide chip design direction.

But with such a long software stack, why start specifically from model research?

The company founder Ren Qingyuan's answer is to start with algorithms and real customer needs, get the closed loop of data, models, agents, and evaluation running, and then use the compute requirements of real tasks to define the chip.

Paimon-0 is the first public result of this route. It belongs to Primotion's Paimon embodied foundation model series and is also the main model that topped multiple leaderboards this time.

Working backward from models to chip design first requires dividing the compute boundaries: which tasks go to the cloud and which must run locally on the robot. This division directly determines the compute load of the on-device chip.

Against the backdrop of the industry widely advocating "device-cloud collaboration" but lacking a clear technical route, Primotion proposed the "local brain + meta-brain" architecture.

Co-founder Ma Zhiheng believes the device side cannot simply be defined as a cerebellum: after the cloud outputs a plan, whether the task can be executed depends on the device side. During robot execution, it needs to follow the task rhythm, continuously adjust actions based on visual, tactile, and force feedback, and coordinate the timing of perception, computation, and control. This kind of high-frequency closed loop cannot rely on frame-by-frame cloud responses and must be completed locally. The cloud handles complex task planning, while the device side must properly carry out each step of execution.

In this system, the cloud meta-brain is responsible for open-scenario understanding and task planning, accumulating reusable task experience, and replanning and failure recovery when tasks get stuck or fail; the device-side local brain is responsible for action execution, relying on multiple types of perception signals for real-time fine-tuning. It can be summarized as: the meta-brain handles the first attempt, and the local brain handles stable repeated execution tens of thousands of times.

Primotion's salad-making demo is exactly what this division of labor looks like. The meta-brain does long-horizon planning and decides which vegetables to pick; the local brain handles grasping and placing, squeezing sauce, mixing, and serving:

With the division of labor in place, the next issue to solve is more practical: when switching to a different "dish," how much cost does the local brain need to learn it?

The most expensive part of teaching a robot to work is data.

Ma Zhiheng has done the math. For a single task, post-training to raise the success rate is often enough to meet customer requirements. What truly bottlenecks things is data collection: a single task often requires collecting 500 to 1,000 data points, and after collecting in the lab, more collection is still needed at the customer site.

Therefore, Primotion is betting that data costs will continue to decline and wants a sufficiently strong cold start: one or two demonstrations first get a new task's success rate to 50% to 60%, and the rest is left to rollout and reinforcement learning to raise further.

To this end, Primotion built an algorithm closed loop comprising data, models, and self-improvement (RSI).

At the data level, model pretraining makes extensive use of first-person human operation videos, mainly supplied by partner EgoScale (Shenzhen Future Eye Technology). EgoScale provides Paimon-0 with hundreds of thousands of hours of first-person operation data. Its collection network covers more than 160 cities in 3 countries and more than 12,800 real scenarios, accumulating more than 4 million tasks, more than 120 million action units, and tens of thousands of "person-scenario" samples.

The dataset balances diversity in scenarios, skills, difficulty, and operating habits; combined with self-developed AI-assisted manual high-precision atomic action annotation and 3D head-hand trajectories, it supports the model's generalization ability and leaderboard performance.

At the same time, Primotion has built its own data processing pipeline for embodied training, converting raw videos into structured action units, completing the mapping from hand trajectories to the target dexterous hand, and performing data quality screening and training adaptation.

Next, at the model level, Paimon-0 is centered on the World Action Model (WAM). Currently, WAM+IDM (inverse dynamics model) is a mainstream route with outstanding results: WAM predicts future states, and IDM outputs actions based on state changes.

However, continuously rolling out complete future videos incurs high compute overhead and latency, making it difficult to meet the real-time operating requirements of robots. Paimon-0 uses query token distillation to extract core representations from future states, including object 3D pose, semantics, and operability affordance, which are then sent to the device-side model for action prediction.

This scheme compresses the inference time of a single action chunk from about 3,000 ms to 300 ms, reducing it to one-tenth of the original; this metric is the action chunk inference time and is not equivalent to the underlying control loop period.

After Paimon-0, Primotion is also training an ego foundation model, likewise pretrained on first-person human operation data, which can already drive a dexterous hand to complete tasks. According to introduction, this ego foundation model is expected to be officially released in 3 months:

Finally, there is the self-improvement module, which co-founder Chang Xinyuan regards as the core breakthrough point for embodied intelligence deployment. Relying solely on large amounts of rollout interaction trial and error to train robots makes large-scale deployment difficult. True self-improvement requires the model to understand the complete execution chain: perceive the current state, reason about the causal impact of actions, and identify the stage of the task, and generalizable automatic evaluation is the foundation of all this. Without automatic evaluation capability, reward signals can only rely on manual annotation, and the structural dilemma of high data costs is difficult to break through.

Around this judgment, the entire self-improvement system consists of three modules working together:

This system relies on a real2sim solution to roll out future trajectories in simulation, integrating information such as 3D position, body pose, and action feedback, and then overlaying the rolled-out states onto the real scenario to help the model intuitively predict action effects and complete decision-making.

Obtaining low-cost, generalizable reward signals is key to continuous robot self-iteration. The Progress Model is used to convert execution results into training feedback.

Primotion's trained Paimon-Eval requires only one demonstration data point for in-context learning to determine the execution stage of an unfamiliar new task. Taking the towel-folding task as an example, when multiple rollouts run in parallel, it can continuously identify which stage execution is in—searching for the target, grasping, folding, aligning, or confirming—and output a task progress curve.

According to Chang Xinyuan, Paimon-Eval is planned for open source in November–December and will be offered in multiple size versions. The team has expanded its training scale to tens of thousands of hours and adopted a task-centric training paradigm, enabling it to recognize task progress and execution quality across different operators and robot embodiments.

On this point, Ren Qingyuan added that algorithms ultimately have to return to deployment. Whether the collected data is effective and can be automatically annotated, how much capability the model retains after quantization and distillation, and its actual performance once deployed on the machine all need to be continuously verified by an evaluation system. Only by stringing these steps into an automated pipeline can the algorithm loop possibly consolidate into a truly deliverable product.

Now that the technical roadmap is clear, let's look at how this team was assembled.

The story of Primotion starts with a chip industry veteran's "head start."

This chip veteran is Ren Qingyuan, CEO of Primotion. He is a science and technology expert at both the Guangdong provincial and Shenzhen municipal levels. During his time as vice president at a leading RISC-V chip company, he led his team to achieve mass production and delivery of chips, with the company's revenue jumping from the tens of millions to hundreds of millions of yuan and cumulative orders exceeding 1 billion yuan.

Ren Qingyuan's previous company developed self-designed RISC-V processors and held core microarchitecture, and he had been looking for a new computing platform. In early 2025, robots entered his field of view, though at the time they seemed capable only of performing; by mid-year he noticed that players were beginning to land deployments one after another, with orders and deliveries following in quick succession.

In Ren Qingyuan's view, this is precisely the window for chip companies to enter:

You can't wait until the boom arrives to start; by then customers are already using someone else's chips.

Thus Primotion came into being. And the algorithm partner he brought in was a pair of fellow students.

Ma Zhiheng and Chang Xinyuan both come from the Institute of Artificial Intelligence and Robotics at Xi'an Jiaotong University, where they studied under Professor Gong Yihong. Gong Yihong is an IEEE Fellow and one of the earliest scholars in China to study deep learning; he once served as head of NEC's Silicon Valley research institute, which produced a large number of AI leaders such as Yu Kai, Lin Yuanqing, and Zhou Xi.

Senior fellow Ma Zhiheng is a distinguished associate professor and doctoral supervisor at the School of Computing Microelectronics at Shenzhen University of Advanced Technology, and also a serial entrepreneur. He has received the World Artificial Intelligence Conference (WAIC) "Yunfan Award," published more than 50 papers at top conferences as a principal author, accumulated over 3,000 Google Scholar citations, served as an area chair for conferences such as NeurIPS and ICML, and has long been engaged in research on computer vision and embodied intelligence.

It is precisely such a scholar who describes Ren Qingyuan as "knowing chips very well and able to quickly grasp algorithms too."

Junior fellow Chang Xinyuan came out of a big tech company; he was previously the technical lead for the embodied intelligence direction at Amap's Visual Technology Center, where he led general embodied projects such as ABot-M, ABot-PhysWorld, and ABot-Claw, with over 2,000 Google Scholar citations. His reason for leaving the big company was very practical: embodied intelligence requires the entire chain—algorithms, data, real machines, and even chips—to iterate together, and model R&D must stay close to edge-side chips.

Between algorithms and chips, someone is still needed to connect the control link, and that is co-founder Zhang Dongjun. He is responsible for domain design, real-time communication, and time-determinism solutions. He has led the R&D and deployment of motion controllers and drone systems at Googol Technology and DJI, is among the earliest batch of doctoral students of robotics authority Professor Li Zexiang, and began researching dexterous hands as early as 2005.

The four each oversee their own area, while Ren Qingyuan oversees strategy at Primotion and also keeps an eye on the compute-control integrated SoC, domain controllers, and the supporting software stack.

In just two months, Primotion expanded to about 50 people. Beyond the algorithm team in the spotlight, there is also a hardware and systems team composed of chip veterans hidden beneath the surface; they are one of the few teams in China to have twice taken large chips from 0 to 1 all the way to mass production of tens of millions or even hundreds of millions of units. The three groups are bound together by a shared engineering culture, deriving SoC architecture and microarchitecture backward from task types.

Incidentally, the name Paimon was coined by the young people on the algorithm team, taken from a character in the game Genshin Impact. Ma Zhiheng said with a smile that the industry has nearly used up all the Greek letters for naming models, so it might as well pick a codename that is both interesting and catchy.

Remember the control timing error reduced to about 1/200 of the original mentioned at the beginning? In fact, many times a robot's problems do not lie in the model at all.

Ren Qingyuan shared a set of experiments: on NVIDIA chips, using the industry's common camera access scheme, from the camera's photosensitive capture to the model reading the image, the process goes through image preprocessing, memory transfer, and synchronization waits, and the entire link takes almost as long as a single model inference. For high-frequency closed-loop robots, even if model inference is fast, if data supply is not timely, actions will still lag.

This is also a key difference between robots and autonomous driving computing architectures: autonomous driving can tolerate a certain delay, whereas robots involve contact-based operations and require higher control frequencies; delay and jitter will directly cause trajectory deviation, unstable grasping, and force-control failure.

Primotion's approach is to mount an FPGA on existing main-control peripherals and offload the preprocessing and postprocessing tasks before and after model inference onto it; most of this RTL is subsequently planned to be integrated into its self-developed chip, and the optimization of control timing error is a typical case of hardware link tuning. As Ren Qingyuan put it:

All engineering problems are, in fact, about cheating under the constraints of existing technology.

But hardware is only part of it, and Ren Qingyuan often says: chip companies are ultimately software companies at their core. In the PC industry chain, firmware and system integration are handled by dedicated vendors, whereas the robotics track lacks such supporting players, so the related work can only be filled in by the chip side. Domestic chips are hard to deploy, and many bottlenecks are stuck in the software stack. Primotion hopes to build a complete end-to-end solution, benchmarked against Nvidia's approach to robotics.

As for the business model, Ren Qingyuan also laid out his plan:

If the customer has nothing, they just buy a software-hardware integrated box, "plug in an Ethernet cable and it works right away"; if they only want the chip, that is sold too. Gross margin comes from the chip, while software is more of an ecosystem issue, so whatever can be open-sourced is open-sourced as much as possible.

At the same time, to make the ROI positive, Primotion's goal is to push the cost of data collection and post-training down to within one-tenth of the robot body, or even to let production line workers collect data and train models themselves until the robot goes online and works.

Having thought through what it wants to do, Primotion has also drawn its own boundaries—

Primotion does not make the robot body, does not compete for profit with model companies or body companies, and does not pursue "one chip for many uses."

Facing the industry's future evolution, Ren Qingyuan very candidly said that the next two years will certainly bring changes, and the key is whether there is a rational judgment framework. In Ren Qingyuan's view, model bottlenecks and chip bottlenecks are mutually causal; some fast motion perception links cannot keep up, and no matter how good the algorithm is, it cannot be put to use.

What Primotion wants to do is first get the end-to-end latency and compute chain right, breaking the deadlock of which came first, the chicken or the egg. According to the plan, its self-developed chip will begin delivery in 2028.

This intention of "giving the underlying layer a push first" was in fact written into the company's name long ago. Primotion's English name, Primotion, comes from Aristotle's concept of the "prime mover," and they want to become that "origin of motion" in the world of physical AI.

And the endgame of embodied intelligence is, after all, an intelligent system that can act stably in the real world; no matter how large the model is, it is only one part of it.

On this road, Primotion has only just begun.

(PS: Primotion is recruiting talent in embodied algorithms, robotic systems, inference quantization, chip design, and other directions. Resumes or cooperation intentions are welcome at hr@primotion.ai.)