By Jin Lei, from QbitAI
QbitAI | Official Account QbitAI
What a contrast.
A company founded in Shenzhenjust three months agoclearly positions itself as makingembodied chips, yet the first result it has delivered to the outside world is a series ofembodied foundation modelsnamedPaimon.
And the first model in this series,Paimon-0, upon its debut, directly claimed four first places:
First, look at the most prestigious one,RoboColiseum, a leaderboard jointly built by nearly 20 global universities, enterprises and research institutions including Hugging Face, NVIDIA and Tsinghua University, which has recently been extremely popular in the embodied AI community; Paimon-0 ranked first with a score of 0.7754.
Note that this leaderboard recently invited general-purpose 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' scores in instruction following and spatial understanding were indeed impressive, on the general manipulation sub-leaderboard, which requires real hands-on action, Paimon-0's 0.6613 was still the highest overall, and it also led the perturbation adaptation sub-leaderboard.
RoboCasa365is a large-scale multi-task benchmark for general robot policies, containing 2500+ simulated kitchen scenes, 365 daily tasks, and 2000+ hours of interaction data. It selects 50 tasks to test generalization ability at three levels: atomic visible, compositional visible, and compositional invisible. Paimon-0 topped it with an overall score of 58.1.
VLA-Arenais an open-source evaluation leaderboard for general embodied models, conducting systematic assessments around safety and disturbance resistance, supporting cross-model comparison. Paimon-0 achieved an average success rate of 0.73, ranking first.
Meanwhile, Paimon-0 participated in theIROS Workshopchallenge organized by Qiyuan Robotics, whose track focused on dual-arm robot policies, including online simulation evaluation and offline real-robot verification. Yanxi Power ultimately won the championship with a total score of 46.599, validating the capability of end-to-end policies in dual-arm manipulation.
And the contrasts don't stop there.
Before today, this company's results had not been made public (there was barely even any news about the company itself), yet its solution had already received customer orders!
Specifically, a customer brought two components of a robotic arm to their laboratory, and after switching to their control solution, a set of on-site acceptance records showed that, at a 1kHz control tempo, the worst-case arrival jitter of the EtherCAT control cycle dropped from 30 microseconds in software mode to 152 nanoseconds in FPGA hardware mode, down to about1/200。
In software mode, the control timing is scheduled by CPU software and is prone to jitter caused by system interrupts and task switching; FPGA hardware mode sinks time-deterministic control down into hardware logic, so the error is greatly narrowed.
Even the customer could hardly believe it at first, but after repeated verification, the facts were just so.
So who exactly is this dark horse from Shenzhen? Without further suspense, it isYanxi Power (Primotion)。
While the results deserve recognition, questions follow:
Why would a chip company first deliver a model report card? Founded only three months ago, how were these achievements achieved?
With these questions in mind, we had an in-depth exchange with the Primotion team.
Why would a chip company start with models?
To answer this question, we must first look at what kind of solution Primotion intends 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 full pipeline of training adaptation, deployment optimization, and production inference.
On the training side, a data processing pipeline 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, a cloud-based meta-brain handles task planning and failure recovery, an edge-side local brain performs policy inference, and an 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 an embedded software stack and compute-control hardware, interfacing with the robot's local control loop.
Primotion wants to connect these links, and around a compute-control integrated SoC and domain controllers, form a hardware-software solution that can be delivered into customers' hands.
This approach has precedent in the autonomous driving field: NVIDIA built system software, middleware, and inference tools around in-vehicle hardware to support upper-layer algorithms landing on chips, with the DRIVE platform being a typical reference. Migrating to the robotics track likewise requires first clarifying algorithm execution, data flow, and task collaboration logic before it can guide the chip's design direction.
But with such a long software stack, why start precisely with model research?
The company's founderRen Qingyuangave the answer: start with algorithms and real customer needs, run through the closed loop of data, models, Agents, and evaluation, and then use the computing demands of real tasks to define the chip.
Paimon-0 is the first public result of this route; it belongs to Yanxi Power's Paimon embodied foundation model series and is the flagship model that topped multiple leaderboards this time.
Deriving chip design backward from the model first requires partitioning compute boundaries: which tasks run in the cloud and which must run locally on the robot; this division of labor directly determines the compute load of the edge-side chip.
At a time when the industry commonly talks about "edge-cloud synergy" but without a clear technical route, Yanxi Power has proposedThe "local brain + meta brain"architecture.
Co-founderMa Zhihengbelieves the edge side cannot simply be defined as the cerebellum: after the cloud outputs a plan, whether the task can actually be carried out depends fundamentally on the edge side. During execution, the robot needs to follow the rhythm of the task, continuously adjusting its motions based on visual, tactile, and force feedback, coordinating the timing of perception, computation, and control. Such high-frequency closed loops cannot rely on frame-by-frame responses from the cloud and must be completed locally. The cloud handles complex task planning, while the edge side must carry out every step of execution well.
In this system, the cloud meta brain is responsible for open-scene understanding, task planning, accumulating reusable task experience, and re-planning and fault recovery when tasks get stuck or fail; the edge local brain is responsible for motion execution, relying on multiple types of perception signals for real-time fine-tuning. It can be summarized as:The meta brain is responsible for the first attempt, and the local brain is responsible for stably repeating the execution tens of thousands of times.
△ Note: the local brain corrects motions in real time through force feedback closed loops.
Video link:
https://mp.weixin.qq.com/s/cSD3nAlAJj1LPH28Qbcb_A
Yanxi Power's salad-making demo is exactly what this division of labor looks like. The meta brain does long-horizon planning, deciding which vegetables to pick; the local brain handles grasping and placing, squeezing sauce, stirring, and serving:
Video link:
https://mp.weixin.qq.com/s/cSD3nAlAJj1LPH28Qbcb_A
Embodied intelligence's low-cost iteration closed loop
With the division of labor settled, a more practical problem remains: switching to a new "dish" — at what cost can the local brain learn it?
When teaching robots to do work, data is the most expensive part.
Ma Zhiheng has done the math. For a single task, raising the success rate through post-training can often already meet customer requirements; the real bottleneck is data collection — a single task often requires collecting 500 to 1000 data entries, and after collecting in the lab, you still have to go to the customer's site to continue collecting.
So Yanxi Power is betting that data costs will keep falling, and it wants a strong enough cold start: achieving a five-to-sixty percent success rate on a new task with just one or two demonstrations, then leaving the rest to rollout and reinforcement learning to push higher.
To this end, Yanxi Power has built an algorithmic closed loop comprisingdata、, models, andself-improvement (RSI).
At the data level, the model's pre-training makes extensive use of first-person human operation videos, mainly supplied by partner EgoScale (Shenzhen Future Eye Technology). EgoScale provides Paimon-0 with over a hundred thousand hours of first-person operation data, with a collection network covering 160+ cities in 3 countries and more than 12800 real scenes, accumulating over 4 million tasks, more than 120 million action units, and tens of thousands of "person-scene" samples.
The dataset balances diversity in scenes, skills, difficulty, and operation habits; combined with in-house AI-assisted high-precision atomic action annotation and head-hand 3D trajectories, it supports the model's generalization ability and leaderboard performance.
At the same time, Yanxi Power has built its own data processing pipeline for embodied training, converting raw videos into structured action units, completing the mapping of hand trajectories to target dexterous hands, data quality filtering, and training adaptation.
Video link:
https://mp.weixin.qq.com/s/cSD3nAlAJj1LPH28Qbcb_A
Secondly, at themodel level, Paimon-0 takes the world action model (WAM) as its core. Currently, WAM+IDM (inverse dynamics model) is the mainstream approach with outstanding results: the WAM predicts future states, and the IDM outputs actions based on state changes.
However, continuously rolling out complete future videos incurs high computational cost and latency, making it difficult to meet the real-time operation requirements of robots. Paimon-0 uses query tokens to distill the core representations of future states, including object 3D poses, semantics, and affordances, which are then fed into an on-device model for action prediction.
This approach compresses the inference time of a single action chunk from about 3000ms to 300ms, reducing it to one-tenth of the original; this metric refers to action chunk inference time and is not equivalent to the underlying control loop cycle.
After Paimon-0, Yanxi Dynamics is also training an ego foundation model, likewise pretrained on egocentric human manipulation data; it can already drive a dexterous hand to complete tasks. According to reports, this ego foundation model is expected to be officially released in 3 months:
Video link:
https://mp.weixin.qq.com/s/cSD3nAlAJj1LPH28Qbcb_A
Finally, theself-improvement module, which co-founder Chang Xinyuan identifies as the core breakthrough point for embodied AI deployment. Training robots purely through large amounts of rollout interaction and trial-and-error makes it hard to achieve deployment at scale. True self-improvement requires the model to understand the complete execution chain: perceiving the current state, reasoning about the causal impact of actions, and identifying the stage of the task — and generalizable automatic evaluation is the foundation of all of this. Without automatic evaluation capability, reward signals can only rely on human annotation, and the structural predicament of high data costs is hard to break through.
Based on this judgment, the entire self-improvement system consists of three modules working in coordination:
- The first is thecloud-based large model, responsible for decomposing tasks, verifying policies, recovering from errors, and producing successful and failed rollouts at low cost;
- The second is the self-developed on-devicePolicy Model(Paimon-0), which improves the success rate and throughput of proficient tasks through reinforcement learning post-training;
- The third is theProgress Model(Paimon-Eval), responsible for judging task progress and locating failure points, providing feedback for subsequent training.
This system relies on a real2sim approach, rolling out future trajectories in simulation, fusing information such as 3D positions, proprioceptive poses, and action feedback, and then overlaying the rolled-out states onto the real scene, helping the model intuitively anticipate action effects and complete decisions.
Video link:
https://mp.weixin.qq.com/s/cSD3nAlAJj1LPH28Qbcb_A
Obtaining low-cost, generalizable reward signals is key to robots' continuous self-iteration, and the Progress Model is used to transform execution results into training feedback.
Paimon-Eval, trained by Yanxi Dynamics, requires only a single demonstration data point for in-context learning to identify the execution phase 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 of execution—finding the target, grasping, folding, aligning, or confirming—is underway, and output a task progress curve.
Video link:
https://mp.weixin.qq.com/s/cSD3nAlAJj1LPH28Qbcb_A
According to Chang Xinyuan, Paimon-Eval is planned to be open-sourced in months 11–12, offering versions in multiple sizes. The team extended 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 operators and across robot embodiments.
To this, Ren Qingyuan added that algorithms ultimately all come back to deployment. Whether the collected data is effective, whether it can be automatically annotated, how much capability the model retains after quantization and distillation, and its actual performance once on the robot—all of these need to be continuously verified by an evaluation system. Only by chaining these steps into an automated pipeline can the algorithmic loop settle into a truly deliverable product.
With the technical route explained clearly, let's now look at how this team was assembled.
A Composite Team of Chips, AI, and Robotics
The story of Yanxi Dynamics must begin with a chip person's"lead time".
That chip person is Ren Qingyuan, CEO of Yanxi Dynamics. He is a science and technology expert at both the Guangdong Provincial and Shenzhen Municipal levels. While serving as Vice President at a leading RISC-V chip company, he led his team to achieve mass-production chip delivery, growing the company's revenue from the tens-of-millions level to hundreds of millions of RMB, with cumulative orders exceeding 1 billion RMB.
The company Ren Qingyuan previously worked at developed its own RISC-V processors with core microarchitecture in hand, and he had been searching for a new computing platform. In early 2025, robots entered his field of vision, although at that time they still seemed to be only for show; it was not until mid-year that he noticed people beginning to land deployments one after another, with contract signings and deliveries coming one after another.
In Ren Qingyuan's view, this is exactly the window for chip companies to enter:
You can't wait until the trend arrives to start working—by then, customers are already using someone else's chips.
Thus, Yanxi Dynamics was born. And the algorithm partners he brought in were two fellow apprentices.
Ma Zhiheng and Chang Xinyuan both come from the Institute of Artificial Intelligence and Robotics at Xi'an Jiaotong University, studying under Professor Gong Yihong. Gong Yihong is an IEEE Fellow, one of the earliest scholars in China to research deep learning, and formerly served as president of NEC Labs Silicon Valley, from which a large group of AI leaders such as Yu Kai, Lin Yuanqing, and Zhou Xi emerged.
The senior apprentice Ma Zhiheng is a Distinguished Associate Professor and doctoral supervisor at the School of Computing Power Microelectronics, Shenzhen University of Advanced Technology, and also a serial entrepreneur. He won the "Yunfan Award" at the World Artificial Intelligence Conference (WAIC), has published more than 50 top-conference papers as a main author with over 3,000 Google Scholar citations, has 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 this scholar who evaluated Ren Qingyuan as "very knowledgeable about chips, and also able to quickly understand algorithms."
The junior apprentice Chang Xinyuan came out of a major tech company, previously serving as the technical lead for embodied intelligence 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 of algorithms, data, physical robots, and even chips to iterate together, and model R&D must stay close to edge-side chips.
Between algorithms and chips, someone was still needed to connect the control link—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, was among Professor Li Zexiang, the doyen of the robotics world's, earliest cohort of doctoral students, and began researching dexterous hands as early as 2005.
The four each manage their own domain, while Ren Qingyuan oversees strategy at Yanxi Dynamics and also keeps an eye on the compute-control integrated SoC, domain controllers, and the accompanying software stack.
In just two months, Yanxi Dynamics grew to around 50 people. Besides the algorithm team standing in the spotlight, hidden beneath the surface is a hardware and systems team composed of chip veterans—they are one of the few teams in China to have twice taken large chips from 0 to 1 and into mass production of tens of millions or even hundreds of millions of units. The three groups are bound together by a shared engineering culture, working backward from task types to SoC architecture and microarchitecture.
By the way, the name Paimon was chosen by young members of the algorithm team, taken from a character in the game Genshin Impact. Ma Zhiheng said with a laugh that the industry is running out of Greek letters for naming models, so they might as well pick a codename that is both interesting and easy to say.
AI Is Important, but It's Not Everything
Remember the control timing error reduced to about 1/200 of the original mentioned at the beginning? In fact, in many cases, the robot's problems don't lie in the model at all.
Ren Qingyuan shared a set of experiments: on NVIDIA chips, using the industry-standard camera access solution, from the camera's photosensitive capture to the model reading the image, the pipeline passes through image preprocessing, memory transfers, and synchronization waits, and the entire chain takes almost as long as a single model inference. For high-frequency closed-loop robots, even if model inference is fast, when data supply is not timely, actions still lag behind.
This is also the key difference between robotics and autonomous driving computing architectures: autonomous driving can tolerate a certain amount of latency, while robots involve contact-based operations with higher control frequency requirements, and latency and jitter directly cause trajectory drift, unstable grasping, and force control failure.
Yanxi Dynamics' solution is to attach an FPGA to the existing main controller's peripherals, offloading the preprocessing and postprocessing tasks before and after model inference onto it; most of this RTL is later planned to be integrated into the self-developed chip, and the optimization of control timing error is a typical case of hardware pipeline tuning. As Ren Qingyuan said:
All engineering problems are essentially cheating within the constraints of existing technology.
But hardware is only one part. Ren Qingyuan often says: chip companies are, in essence, software companies in the end. In the PC industry chain, firmware and system integration have dedicated vendors to take on the work, while the robotics track lacks such supporting players, and the chip vendor has to fill in the related work. The difficulty of deploying domestic chips largely stems from bottlenecks in the software stack. Yanxi Dynamics hopes to build a complete end-to-end solution, benchmarking NVIDIA's approach for robotics.
As forthe business model, Ren Qingyuan also presented his solution:
If a customer has nothing, they can buy an all-in-one software-hardware box — "just plug in a network cable and it works right away"; if they only want the chip, that is sold too. The margin comes from the chip, while software is more of an ecosystem issue, so as much as possible will be open-sourced.
At the same time, to make the ROI work out, Yanxi Power aims to push the costs of data collection and post-training down to less than one-tenth of the robot body's cost, even enabling production line workers themselves to collect data and train models until the robot is up and running.
Having figured out what to do, Yanxi Power has also drawn its own boundaries —
Yanxi Power does not make robot bodies, does not compete with model companies or robot-body companies for profit, and does not pursue "one chip for multiple uses".
Facing the industry's future evolution, Ren Qingyuan frankly stated that changes will certainly come within the next two years, and the key lies in whether one has a rational judgment framework. In Ren Qingyuan's view, the bottleneck of models and the bottleneck of chips are mutually causal: in some fast-motion scenarios, the perception pipeline cannot keep up, and no matter how good the algorithm is, it cannot be put to use.
What Yanxi Power wants to do is first get the end-to-end latency and computing pipeline right, breaking the chicken-and-egg deadlock. According to plan, their self-developed chip will begin delivery in 2028.
This intention to "push from the underlying layer first" was actually written into the company's name from the start. Yanxi Power's English name, Primotion, is taken from Aristotle's concept of the "prime mover" — they want to be that "origin of motion" in the world of physical AI.
And the endgame of embodied intelligence is, after all, intelligent systems capable of acting stably in the real world; no matter how large the model, it is only one part of it.
On this road, Yanxi Power has only just begun.
(PS: Yanxi Power is recruiting talent in embodied algorithms, robotic systems, inference quantization, chip design and other areas. Resumes and partnership inquiries are welcome at hr@primotion.ai.)