A Ph.D. student at Tsinghua University, born after 2000, founded OriginFlow, a company specializing in neuromuscular signal data collection, which has raised over 500 million yuan in funding.

Embodied intelligence data collection has added a new pathway.

Before we get into the official introduction, let’s do a quick interactive exercise: Without relying on AI or looking up information, can you name the existing embodied data collection methods within 5 seconds?

The four most mainstream methods are real-device remote operation collection, body-less portable collection (UMI/Ego), simulated synthetic data, and internet video distillation.

(Ranked in no particular order—I’m just listing them off the top of my head)

If further categorized, an extensive list could be compiled.

At least the answer AI gave me is like this—everyone can feel it—

Recently, a new approach has emerged and gained popularity, shifting the focus of data collection from external actions to the body's movement output.

This is the currently widely discussed surface electromyography (sEMG) solution.

OriginFlow (Yuanche Taichu) is the most popular and highly sought-after startup on this path.

Founder and CEO Qin Shentao, born in 2001, is currently pursuing his Ph.D. at the School of Vehicle and Mobility at Tsinghua University.

In August last year, he founded OriginFlow in Beijing; as of May this year, the company has publicly disclosed that it has successfully completed its angel, strategic, and Pre-A1 funding rounds, raising a total of over RMB 500 million, with investors including BlueChilli, Oasis, and Monolith.

Qin Shentao told us:

There is a missing layer that can abstract non-structured, high-precision, and complex physical interaction information from real-world industrial scenarios into trainable representations.

Therefore, he founded a startup focused on physical AGI infrastructure, aiming to use neuromyoelectric signals as one of the entry points for embodied data collection.

Qin Shentao said that, in the short term, solutions like UMI and Ego’s first-person vision represent incremental improvements for NeuroScale; however, seamless data collection will undoubtedly be a significant trend in the future.

When new innovations emerge, a more suitable strategy is to evolve alongside existing technologies.

Let’s first get a practical sense of what OriginFlow is doing right now.

When a person tightens a screw, steadily holds a bowl of soup, or grips a soft object, the body continuously provides subtle feedback based on the situation.

When do the fingers start to tighten? How is the pressure increased or decreased?… In reality, people don’t consciously think through each of these questions—one’s body quickly and automatically makes and executes a series of judgments regarding contact, friction, weight, and stability.

OriginFlow aims to record the reactions involved in this layer as part of embodied intelligence data.

To understand why we bet on this direction, we must go back to Qin Shentao's undergraduate years.

During his undergraduate years, Qin Shentao led his team to win nearly every robotics competition championship available; it was also during this time that he began to think more deeply:

Keyboards, mice, and voice are not the most direct ways for humans to communicate with machines. Human hands are the most flexible tools; if machines can understand the true intentions behind hand gestures, human-machine interaction will reach a new level.

Thus, neural interfaces came into Qin Shentao’s field of vision—he is deeply committed to human-machine integration.

Neural interfaces capture surface electrical signals related to peripheral motor nerves (i.e., neuroelectromyographic signals), avoiding the challenges of high invasiveness, high channel counts, and brain region mapping associated with brain-computer interfaces.

Since its official inception, electromyography has been confined to university laboratories.

Later, with advancements in wireless sensing, dry electrodes, and AI algorithms, neuromyoelectric signals began to be used as an entry point for perceiving human movement intent, and are now widely applied in biofeedback training for rehabilitation medicine, athletic movement analysis, occupational ergonomics and injury assessment, and discrete motion control of myoelectric prosthetics.

△

Here’s a typical example~

In 2019, Meta’s core research division, Reality Labs, acquired a non-invasive neural interface company called CTRL-Labs.

Before the acquisition, CTRL-Labs had already released a developer kit featuring a main device the size of a watch and a tethered component containing electrodes.

It uses a 16-channel EMG sensor to detect muscle electrical signals transmitted from motor neurons to the fingers at the wrist, and employs AI algorithms to decode these signals into digital commands (such as clicks, swipes, and gestures).

Two years later, Meta publicly demonstrated a prototype of a neural interface wristband based on CTRL-Labs technology, called the Meta Neural Band, showcasing its potential to control AR interfaces by capturing subtle neural signals—for example, typing in the air or selecting menus.

At the time, Xiao Zha said that this interaction method was "nearly infinite in control" because it could precisely sense users' intentions—even for actions they had not yet made.

Last year, Meta's third-generation smart glasses introduced a neural input wristband, allowing users to directly control the glasses' interface through hand gestures—such as imagining swipes and clicks.

(An off-chain note: The Apple Watch's pinch gesture for confirmation, closing windows, etc., primarily relies on optical sensors.)

Unlike Meta, which is concerned with making it easier for people to click, swipe, or type, Qin Shentao’s thinking focuses on the missing conditions for physical AGI.

In his view, the severe shortage of high-quality physical interaction data has become a key bottleneck in the advancement of embodied intelligent robots.

Current methods of embodied data collection each have their drawbacks—could electromyography serve as a complementary approach?

Moreover, selecting neuromuscular electrical signal acquisition may bypass differences in contact surface materials and sensors, enabling the identification of a shared action representation that connects muscle activation, tendon force, and joint motion from the perspective of both human and machine drivers.

In 2025, Qin Shentao, a Ph.D. candidate, officially founded OriginFlow with the vision of building an "action foundation" for Physical AGI beyond text and video.

Soon, the team introduced the NeuroScale technology framework.

NeuroScale is not simply adding an EMG wristband to existing embodied data collection devices.

It is a data and model framework encompassing signal acquisition, physical quantity reconstruction, action representation, and cross-ontology transfer.

This system uses a non-invasive neuro-motor interface as the signal entry point, integrating multimodal data such as electromyography, first-person visual input, and IMU sensors. Through the PULSE base model, it reconstructs posture, contact forces, and driving forces during human operations, and encapsulates each real-world operation into machine-learnable Human Tokens.

NeuroScale has long focused on two core issues.

First, how can Human Data Scale Up be achieved?

By continuously recording genuine user actions with minimal disruption to natural perception and movement, we generate more high-quality physical interaction data.

Second, how to achieve cross-embodiment transfer between humans and robots, enabling human movement experience to be represented and adapted for robots with different physical structures.

Beneath these two questions lies a more fundamental technical issue: In what form should human actions be fed into the base model?

The text has developed a relatively mature token representation, and visual modalities have gradually converged toward patch or latent representations, but the action modality still lacks a widely accepted standardized representation.

OriginFlow breaks down a single physical action into three interconnected spaces:

MotionSpace describing hand gestures and movement trajectories

Describe TactileSpace for normal force, tangential force, and contact feedback.

Describe TendonSpace for muscles, tendon forces, and joint torques.

Driving force generates motion, motion leads to contact, and contact ultimately produces force—these three elements together form a physical causal chain.

The Human Tokens mentioned by OriginFlow are action representations built upon these three physical quantities.

“In form, humans can be regarded as a special case within the configuration space of embodiments. Therefore, the transition from human to robot is essentially a subproblem of ‘cross-embodiment transfer,’” said Qin Shentao.

Specifically, large-scale Human Data covers a broad spectrum of human movements and skills, while smaller-scale but highly diverse Cross-Embodiment Data provides alignment anchors between humans and various robotic embodiments.

After combining the two types of data, the model can learn a shared action representation that is decoupled from specific body configurations, enabling the redirection of the same human motion and force information to robotic bodies with different degrees of freedom and actuation methods.

Therefore, OriginFlow does not aim to directly replicate human operational data to robots; instead, it first identifies shareable actions, driving forces, and contact relationships among different entities before performing adaptation and mapping.

In the real world, NeuroScale begins with a wristband.

The wristband is named OriginKitGen 1.0, slightly wider than Apple Watch bands but with a smaller overall size and lighter weight.

The wristband captures microvolt-level surface electromyographic signals emitted by the wearer, using a 16-channel acquisition system with an information data stream of approximately 96 KB per second, enabling continuous modeling of hand movements.

However, NeuroScale's implementation doesn't rely solely on the wristband.

With just the wristband, the system can detect changes in forearm muscle activity, but it struggles to determine exactly what the fingers are doing.

Signals captured by OriginKitGen 1.0 are fed into the NeuroScale data pipeline alongside first-person visual and IMU data.

After alignment, calibration, and processing, the original waveforms are analyzed by PULSE, a proprietary base model developed in-house, to extract cues related to hand posture, motion trajectories, contact forces, and tendon actuation—corresponding to three core physical domains: Motion Space, Tactile Space, and Tendon Space. On the hardware side, filtering, differentiation, and motion artifact suppression are applied; on the model side, neural signal encoding and strong supervised learning are performed to progressively transform the data into action representations learnable by machines.

This is what the team calls Human Tokens.

During WAIC, Qin Shentao and his team showcased the current achievements: a demo of the PULSE 0.2 version.

The wristband captures 16-channel surface electromyographic signals, and PULSE uses these signals as input to continuously model hand movements and observe changes in muscle activation during finger opposition.

Unlike discrete gesture recognition, PULSE focuses on continuous hand tracking and fingertip force representation.

When the user performs a finger gesture, the system can observe changes in force in real time.

However, reconstructing an action from a neural signal does not equate to obtaining data that can be directly used for robot training.

The raw data still requires multi-device clock alignment, individual calibration, motion artifact removal, task segmentation, action and force annotation, quality filtering, and cross-modality mapping.

To this end, OriginFlow has also built a Data Infra that covers data production, processing, evaluation, and training.

Among these, ORACLE uses a multimodal model to automatically annotate actions, forces, semantics, and task segments; the unified multimodal foundation CHORD handles temporal and representational alignment between electromyography, vision, IMU, language, and robot states.

The data will also pass through a series of quality checkpoints, including physical validity, task validity, annotation validity, model value, and real-world evaluation. Only data that generates tangible value in model training or robotic tasks will be included in the final delivered dataset.

In other words, it’s not enough to simply collect data—you must also be able to process vast amounts of raw signals into trainable, evaluable, and reusable data assets at a sufficiently low marginal cost.

PULSE 0.2 represents only a small portion of the company's operations, and the model itself is merely an interim output.

PULSE 0.3, currently under development, is further exploring the relationship between tendon force and hand forward kinematics.

The wristband mentioned earlier is merely the current collection terminal, not the final product form.

For a route that relies on long-term, large-scale data collection, whether the equipment is naturally accepted by people determines whether data can be continuously generated.

Qin Shentao said that people must first be willing to wear the device and then be willing to wear it long enough.

At its core, OriginFlow aims to continuously record human movements in the real world while minimizing disruption to existing perception and motion patterns, and transform them into Physical Tokens.

Only then can data collection move beyond data collection facilities and dedicated workstations into daily life and real production processes.

By then, the “trillion-hour data” currently out of reach in the field of embodied AI would require only about ten days of recording the behaviors of all humans.

Qin Shentao described this process as the continuous scaling of data.

Looking ahead, we will scale full-modal Human Data from tens of millions of hours to hundreds of millions, and eventually toward trillions of hours.

The next-generation physical entry points envisioned by the team will gradually become lighter, more natural, and more akin to consumer products.

In their vision, the future will allow people to continuously model themselves in daily life simply by wearing glasses, watches, or lighter wristbands with consumption features, gaining convenience from their personal models.

△

Overall, OriginFlow is currently in the first phase of its strategic development plan.

They have planned a three-phase development roadmap—Phase One, From Human: capture as much as possible about how humans move, systematically distilling and converting human operational experience in the physical world into machine-learnable data and representations.

OriginFlow aims to fully replicate human movement output without interfering with the natural perception and motion processes of humans.

The complete data collection solution always adheres to a non-intrusive principle, ensuring that movements are accurately observed and reconstructed while the person maintains their natural state.

Phase Two: With Human

This phase will develop the next generation of AI hardware, including robots, designed to seamlessly integrate into daily work and life in an always-on form, enabling more people to benefit from the value of the next-generation human-machine interface.

The third stage is Enhance Human, targeting the next 10 to 30 years.

Of course, there is still a long road of engineering validation between a single-point demonstration of PULSE 0.2 and trillion-hour-scale multimodal human data.

It is reported that domestic startups such as BrainCo, Octopus Power, SnowOrigin, Nianxiang Technology, WuJie Maluo, and Shouyi Technology are also using electromyography to collect embodied intelligence data.

This article is from the WeChat public account "Quantum Bit" (ID: QbitAI), authored by Heng Yu.