Xingyuan Intelligence's Yang Shiqi: We Don't Build Robot Bodies, We Bet on Robot "Tokens" and the Cross-Embodiment Data Engine Chorus

In the high-profile entrepreneurial wave of embodied intelligence, most players are rushing into an "arms race" of body manufacturing and large-model architecture. However, Yang Shiqi, a post-2000s scholar who once served as a researcher at a top North American robotics laboratory, holds the opposite judgment: the industry's real crux lies not in hardware, nor in the blind pursuit of model innovation, but in how to achieve large-scale supply and reuse of training data.

After founding Xingyuan Intelligence in June this year, Yang Shiqi did not rush to release a complete robot or preach a grand narrative about models. It was not until October 27 that Xingyuan Intelligence will launch its cross-embodiment data engine "Chorus Zhongxing."

Chorus attempts to clean and perform 3D reconstruction on multi-source human operation videos from smart glasses, chest-worn cameras, UMI devices and others, so that a single human demonstration can be adapted across embodiments to multiple robots, attempting to build the underlying infrastructure connecting human behavior to robot actions.

A week before Chorus's release, he not only dissected the "Demo myth" in the embodied intelligence industry, but also pointed out that the industry is still in the "non-standard automation" stage. Through Xingyuan Intelligence's strategy of insisting on "not making hardware bodies," Yang Shiqi explained the underlying logic of the team's technical path of System Scaling, connecting data, models and real-world deployment to explore the scaled development of robot foundation models.

Post-2000s Yang Shiqi has entered the embodied intelligence track. He does not make bodies, but instead bets on robot "tokens" for his startup.

Diagnosing the embodied intelligence industry before starting up

After returning to China in July 2025, Yang Shiqi spent several months conducting in-depth research on the real progress of embodied intelligence in China, discovering some differences between research and industry.

At the top North American laboratory, Yang Shiqi was one of the world's earliest researchers to use human data for robot training. In his understanding, the correct research path should be that when the first model or product is released, one can already calculate how the next generation of data will be collected, when it will come to market, and how to achieve scaled expansion. By contrast, he saw that some teams in the industry are more keen to chase hot topics in large-model architecture, frequently releasing vision-language-action models (VLA) or world models (World Model), but often lack a clear path for how the system will continue to iterate and how capabilities will scale.

"The embodied intelligence industry generally emphasizes reflecting robot learning capabilities in the form of demos or papers, and even the real model capabilities behind some demonstrations need to be questioned," Yang Shiqi said.

Based on vigilance against "demo prosperity," Xingyuan Intelligence chose to break out of the mindset of complete-machine manufacturing.

Just as the foundation of large language models (LLM) is "tokens," what robot foundation models need to solve is intelligence in the "action" dimension. The name Xingyuan also comes from this. Yang Shiqi hopes to further explore and define robots' "Action Token" in the future.

"Large language models have already made breakthroughs in understanding, memory and planning, but robots must ultimately translate understanding into actions," Yang Shiqi said. What his team wants to do is start from human data and build a model and data system, rather than build another robot.

Expanding human data "40-fold"

The shortage of robot training data is a long-standing pain point in the industry. Early data collection relied on teleoperation, which was costly and limited in scenarios; later solutions such as UMI, although free from the constraints of robot bodies, still failed to fully activate the massive, continuously generated human behavior data in the real world.

"Humans themselves are an embodied form, with rich movement experience and dexterous manipulation capabilities," Yang Shiqi believes. Directly using human videos is an inevitable direction for breaking the data bottleneck. But just because humans can do something does not mean robots can directly copy it. Human videos have problems such as hand occlusion, viewpoint changes and incomplete trajectories, and there are differences in body structure and movement patterns. If unprocessed, the vast majority of data would be directly eliminated in quality screening.

The core of the Chorus engine launched by Xingyuan Intelligence is to solve the problem of data scaling in robot learning: how to turn limited human experience into robot training data that can be reused at large scale.

To this end, Chorus uses atomic action recognition, semantic annotation, 3D motion reconstruction, trajectory optimization and cross-embodiment mapping to bring human operation data from different sources into a unified processing pipeline, enabling the same human experience to be adapted to different robot bodies, thereby expanding the supply scale of effective training data.

Under the current test configuration, Chorus has achieved an effective data conversion rate of about 85%. By adapting to 16 robot bodies, 1 hour of pre-screened human video can generate about 13.6 hours of cross-embodiment action trajectories; further combined with 3x instruction augmentation, it ultimately achieves about 40-fold expansion of effective training data.

However, Yang Shiqi does not believe that data scaling simply means adding more training data. The scaling of data and models needs coordinated design. Only by matching the scale and quality of data with the training needs of models can the Scaling Law in robot learning be further explored.

Of course, whether model capabilities continue to improve as data scale expands ultimately still needs to be verified in real scenarios. This is also why Xingyuan Intelligence emphasizes System Scaling: starting from data scaling, model training and real-world scenario deployment must feed back into each other, continuously driving iterative upgrades of robot capabilities.

Not making bodies, so models can run faster

In terms of strategic choice, Xingyuan Intelligence has drawn a clear boundary: resolutely not making hardware bodies.

“China already has a mature and massive robotics hardware supply chain.” Yang Shiqi explained that building one's own body not only requires huge capital investment but may also slow down the iteration speed of software models. More importantly, what industrial customers need is not a humanoid robot with a complete appearance, but a system that can stably complete specific processes.

This strategic positioning determines Xingyuan Intelligence's technical path: do not make bodies, and do not bind to a single hardware.

Xingyuan Intelligence's solution is that models do not have to wait until they have near-perfect autonomous capabilities before entering practical application. Once a model reaches a certain usable level, it can be deployed in real scenarios; when it encounters situations it cannot handle, humans intervene promptly to complete corrections, and then feed these experiences back into the training system.

He repeatedly emphasized two things: how many times data can be reused, and how quickly the model can iterate in one round.

“A large number of factory customers simply do not care what model or algorithm you use; they only care whether you can get the work done.” Yang Shiqi said bluntly.

In real factory environments, robots are highly prone to out-of-distribution (OOD) situations not covered by training data, causing task interruptions. To this end, Xingyuan Intelligence designed a human-machine collaboration mechanism: when the robot gets stuck, ordinary workers who do not understand technology can quickly take over within 2 to 3 seconds through simple devices, and after correction, return it to the system to continue execution, while the feedback formed by manual takeover continuously feeds back into the model.

In this way, Xingyuan Intelligence tries to strike a balance between task success rate and labor cost. In scenarios such as logistics, initially one worker can supervise one robot (“one-to-one”), and as the model success rate improves, it evolves toward “one-to-three.”

For Xingyuan Intelligence, this is also the difference between System Scaling and simply expanding the scale of models or data. The improvement of robot capabilities depends not only on how much data is invested in the training stage, but also on whether the entire system can continue to iterate in real applications.

The next 3 years are a technical window period for embodied intelligence

Speaking of the current embodied track, Yang Shiqi does not avoid the fact that the industry is in a stage of capital boom and marketing competition, but as a startup newcomer, what he sees is that embodied intelligence still has enough large technical space.

He believes that at this stage the industry should work together to put technology into practice, rather than competing with each other for nominal leading positions.

Regarding the current state of industry development, Yang Shiqi gives an objective judgment: embodied intelligence is essentially still in the “non-standard automation” stage, and it may still take three to four years of accumulation before robots truly broadly assume general labor.

Of course, this is also a technical window period for every embodied intelligence company to achieve breakthroughs.

The embodied industry does not lack imagination about the future, nor does it lack stunning demos, but from a robot completing one task to continuous learning, there is still a long way to go. For Yang Shiqi and Xingyuan Intelligence, Chorus is the starting point of this exploration: starting from human experience, enabling data to be reused at scale, and allowing models to continuously iterate in training and real deployment. What they expect is not only that robots can complete more tasks, but that machine intelligence gradually gains the ability to grow autonomously. (This article was first published on TMTPost APP, written by Qian Fang)