AINEWS
OriginFlow's wristband reads the electrical signals in your muscles and turns them into a single recorded human motion for robot training, a fifth data lane its founder says sits alongside a person driving a robot remotely while it records,
The bottleneck in embodied AI is not a smarter model. It is the data those models learn from. Robots that learn from physical interaction need real, high-resolution, multimodal recordings of how humans actually move through the world, and that kind of data is scarce, expensive to label, and almost impossible to collect at the scale language models take for granted.
A post-2000-born Tsinghua PhD student named Qin Shentao is betting that the faint electrical signals your nerves already send to your muscles can become a fifth mainstream lane in that pipeline. His Beijing startup, OriginFlow (渊澈太初), founded in August 2025, says it has closed more than 500 million yuan, roughly $70 million at recent rates, across angel, strategic, and Pre-A1 rounds as of May 2026, with backers including BlueRun Ventures (蓝驰), Oasis (绿洲), and Monolith. The wager is that a wristband you can wear all day can collect robot training data in the background of normal activity, in a way teleoperation, portable rigs, simulation, and internet video distillation cannot.
The signal is called surface electromyography, or sEMG. It is the same class of faint electrical activity that a physical therapy clinic reads off your skin to gauge muscle recovery. OriginFlow wants to repurpose that signal as a robot training input. Its stack, branded NeuroScale, pairs an sEMG wristband with a first-person camera and an inertial measurement unit, then runs the combined stream through a base model called PULSE. The output is a sequence of "Human Tokens," OriginFlow's coined term for a tokenized unit that captures one real human operation: the muscle activation, the eye-level view, the wrist motion, in a form that can be mapped onto different robot bodies. "Human Tokens" is the company's own framing; it is the unit OriginFlow wants to plug into a larger Physical AGI pipeline.
The bet sits against four other lanes the embodied-AI field has been chasing. Real-robot teleoperation pairs a person with a robot and records both. Portable rigs like UMI and Ego Labs strap a gripper and a camera to a hand and let the wearer go about their day. Simulation generates synthetic trajectories at scale. Internet video distillation scrapes hours of YouTube and tries to extract usable physical knowledge from it. Each lane has tradeoffs: teleop is high-quality but expensive, portable rigs scale better but cap at coarse manipulations, simulation drifts from reality, and video is noisy. Qin frames sEMG as "incremental" against UMI and Ego in the short term, but argues the lane is structurally different because the signals can be collected almost invisibly, the same way a fitness tracker logs steps. He has been public about this framing and about the thesis that invisible data collection is the next embodied-AI substrate.
The technical pitch behind that thesis is that sEMG bypasses some of the surface and sensor differences that plague wrist-worn data. Muscle activation, tendon force, and joint motion share a common substrate, the argument goes, and that substrate can be lifted into a shared action representation that a base model can learn from. That is a founder- and company-stated hypothesis; there is no public benchmark, peer review, or independent comparison yet. The product itself, pricing, channel, and form factor beyond the prototype, is also undisclosed.
Qin is a native of Jincheng, Shanxi, born in 2001. He did his undergrad at Harbin Institute of Technology's School of Mechanical and Electrical Engineering, where his teams won multiple robotics competitions. He surfaced publicly in 2023 at the MiraclePlus (奇绩创坛) spring demo day, then as a serial-founder in the making. He is now a PhD student at Tsinghua's School of Vehicle and Mobility. The 2023 demo day recap captures the pre-OriginFlow chapter.
Meta acquired CTRL-Labs, the sEMG wrist startup, in 2019, and has since demoed a Neural Band prototype and shipped a neural-input wristband alongside its third-generation smart glasses. The Meta timeline is a public fact, but it is also a comparison OriginFlow has not invited: the company has not named Meta as a competitor, has not disclosed pricing or productization, and has not claimed a customer. The honest read is that OriginFlow is the most capitalized bet on sEMG-as-data-substrate that has been publicly disclosed in China, and that it is still a bet, not a shipped product with independent validation.
What to watch next is whether PULSE produces a public benchmark, whether the wristband ships outside a founder-controlled pilot, and whether any of the four incumbent lanes is meaningfully augmented, or displaced, by muscle signals at scale.