Back to Human Data

Human Data

An Interview with Ruoli Dai: Noitom Robotics and the Human-Data Challenge, Part Two

Ruoli Dai explains motion-capture technology, optical and inertial tracking, human-centered data collection and a reusable, cross-embodiment data business.

Zha Dami2025-12-0218 min readWeChat original
Contents

As embodied intelligence entered 2025, certain concepts attracted renewed attention: world models, for example, which we discussed in depth in our 2023 and 2024 interviews—Robot Data, Episode 8: Chuang Gan and Xian Zhou on RoboGen, Generative Models and Differentiable Simulation (Including a Review of Differentiable Physics Engines), and Conversation with Yunzhu Li: Learned Dynamics for General Manipulation—and UMI, covered in Robot Data, Episode 2: Cheng Chi and Zhenjia Xu on UMI, a Physical API for Robot Training. Previously fashionable concepts, such as VLA and teleoperation with matched embodiments, were sometimes pushed aside in these comparisons. Investment in China's embodied-intelligence sector seemed to have become investment in labels. Investors busily searched for projects carrying a particular conceptual tag, while startups busily attached new tags to themselves.

This concept-driven investing closely resembles the earlier Copy to China model. When an overseas company reveals a technical direction, investors rush to find a domestic equivalent. It has also prompted many outstanding Chinese researchers in the United States to found their companies in deep secrecy, fearing that peers in China will copy their work.

The real contribution of embodied intelligence is not the valuations inflated by successive waves of fashionable concepts. Writing in 2025, as a robotics investor a little fatigued by the repetition, I am pleasantly surprised that some domestic entrepreneurs still dare to lead rather than follow. The Human Data series aims to draw attention to a data route that has not yet been overexposed but may hold tremendous potential. I also hope readers notice people willing to take responsibility for creating value and moving the industry forward, rather than following trends merely to raise money.

For the final interview in the Human Data series, I am honored to speak with Ruoli Dai, founder of Noitom Robotics. This is the concluding conversation of our closing interviews. If you are a devoted Smart Note reader, please read it through to the end.

Search for ReSpark on Xiaoyuzhou or Apple Podcasts to listen.

Podcast: Conversation with Ruoli Dai: Motion-Capture Technology and Human Data—Noitom Robotics' Data Story, Part Two. Listen in the original WeChat article.

Original article outline

Contents

↓

1. Introduction

2. How Recent Research Projects Relate to Noitom Robotics

3. An Introduction to Motion Capture

4. Differences Between Motion-Capture Approaches

5. Choosing a Motion-Capture System for Robotics

6. Why Choose Noitom Robotics' Equipment and Data?

7. Near-Term Contributions to the Community

8. Talent Needs

9. Business Partnerships

1. Introduction

I am Ruoli Dai, founder and chief executive officer of Noitom Robotics. We focus on applying motion capture, high-precision tracking and computer vision to support the entire robotics data pipeline. Our work covers collection equipment, data-collection services, post-processing, annotation and licensing our own datasets. We aim to provide complete data infrastructure and technical support for the robotics industry.

We founded Noitom Robotics in 2025. Before that, I founded a motion-capture equipment company called Noitom. The name is the word motion spelled backward. Its core technologies and products revolve around one goal: digitizing human movement.

We have long studied how wearable sensors and computer vision can precisely capture and reconstruct human motion digitally. Over more than a decade—thirteen years, to be precise—we served industries and customers around the world, including film visual effects, virtual reality, human–computer interaction, sports analysis and healthcare. In the relatively specialized inertial motion-capture market, we became a global leader, with a market share exceeding 70% by our estimate.

My new company, Noitom Robotics, carries forward our experience in hardware, sensors and algorithms, applying it to embodied intelligence and robotics. We want to use our capabilities in digitizing human motion, high-precision tracking and motion capture to support this emerging industry with data and technology.

There is a natural logic behind this direction. Humanoid robots and embodied-intelligence systems depend heavily on digitized human interaction, behavior and poses. They need to learn skills and capabilities from those digital records. That underlying logic creates a new opportunity for us and brings me back to robotics.

My PhD research concerned active vibration damping and grasp control in robotics. Returning to the field fifteen or sixteen years after graduating feels both like a continuation and a homecoming.

2. How Recent Research Projects Relate to Noitom Robotics

Yanjie Ze's first TWIST project—see Human Data, Episode 1—MoCap: TWIST—uses optical tracking, probably OptiTrack if my understanding is correct. When I met him at Stanford about a month before this interview, I saw early demonstrations of TWIST2. He was using two PICO controllers, a headset and two foot trackers. Coincidentally, that same day we saw a project released by Westlake Robotics that used our full-body PNS motion-capture system. The video was striking. Yanjie joked that he felt some pressure and wanted to improve his project before releasing it. He had also received a complete set of our whole-body capture equipment and was experimenting with using it for real-time operation. The original transcript suggests adding an interview index here, but provides no specific link.

We also witnessed the XPENG launch that had generated online discussion in the days before this interview. Since late 2023, our technical team had worked closely with XPENG's humanoid-robot team. It is a very capable team, among China's leading groups in hardware manufacturing, motion control, large models, multimodal-model exploration and product development. They used our data and equipment support.

The data behind the female humanoid’s “perfect gait” in that demonstration was collected with our motion-capture equipment and used for imitation learning, reinforcement learning and model training. Confidentiality agreements prevent me from sharing further details.

The engineer operating the robot in Unitree’s recent “out-of-body avatar” teleoperation demonstration and Westlake Robotics' demonstration a few weeks earlier was also wearing our inertial capture equipment. It is called Perception Neuron Studio, which we describe as the world's best-selling motion-capture system. It has seventeen body sensors: seven for the lower body and ten for the upper body. The gloves add ten more, for twenty-seven IMUs in total. These sensors collect raw gyroscope, magnetometer and accelerometer information. Processing at 480 frames per second yields relative positions and absolute orientations for the body segments.

This data enters Axis Studio, which performs three layers of computation. The first is forward kinematics, establishing the overall body configuration. The second is inverse kinematics, helping maintain stable end-effector positions and foot contacts consistent with physical constraints. The third is a physics-engine-based constraint layer, using physical laws and anatomical structure as boundary conditions for optimization. Editorial correction: the Chinese transcript calls the first two layers forward and inverse dynamics, but gives the English terms Forward Kinematics and Inverse Kinematics; those kinematic terms are used here.

The human body has many constraints. We spent considerable time studying its mechanical structure to understand how to digitize movement, even studying anatomy through donated cadavers. The elbow has little tolerance for bending backward; the second finger joint has one degree of freedom, while the first has two; and the radius and ulna allow the forearm to rotate about its axis. Physical constraints include gravity, ground contact, friction, impacts and potential-energy conversion during a jump. These layers of computation transform raw IMU data into digital sequences of human poses. That entire process of digitizing human movement is motion capture.

The projects mentioned earlier use different capture technologies and training methods to achieve results such as XPENG’s “perfect gait”, real-time whole-body teleoperation from Unitree, Westlake Robotics and TWIST, and sparse-tracking-based teleoperation in TWIST2. We communicate closely with these researchers and frequently exchange industry and academic information.

3. An Introduction to Motion Capture

Motion capture means digitizing human movements and poses. The definition does not prescribe a specific method. Computer vision, sensors, magnetic tracking and exoskeletons can all be used. If a method captures human movement and turns it into a digital sequence, it is motion capture.

The technology has a long history. The interview traces Vicon's optical motion-capture lineage back to the 1970s. The British company is a pioneer and benchmark in the field, and I still regard it as setting the standard for optical-capture quality. Its first-generation systems, described in the interview as dating back roughly fifty years, used enormous cameras and tracked only a few markers. Editorial clarification: Vicon's official history dates its first product to 1979 and the start of trading to 1984. Vicon is the motion-capture business of the publicly listed Oxford Metrics group; it is not itself a separately listed company.

James Cameron's first Avatar film helped bring motion capture to a broad public audience and into mainstream production workflows. To achieve real-time visualization and rendering, the production built an entire virtual-production pipeline. Its success won several Academy Awards and helped foster a number of well-known companies providing visual-effects workflow technology.

Noitom was founded in 2012 and introduced its first-generation product in 2014. We adopted a highly disruptive pricing strategy, reducing the cost by roughly two orders of magnitude. Motion-capture work that previously required more than a million yuan could be done for tens of thousands. This quickly attracted users worldwide.

Our equipment has been used widely. Starting with season five of Game of Thrones, for example, group characters such as the White Walkers used our motion-capture equipment. More than ten systems on set allowed seven or eight stunt performers to capture all the movements. Our equipment also participated in visual-effects production for Wolverine, Star Trek, Star Wars and other projects.

In sports, the coaching teams of golfers such as Tiger Woods and Rory McIlroy have long used our equipment to analyze posture and the kinetic chain of the swing. Several Chinese Olympic champions and the national modern-pentathlon team also use our system for digital analysis in routine training.

Our motion-capture technology is also used in virtual reality, games and interaction devices. Companies including Oculus, HTC and Sony PlayStation VR have used our technology licenses, services or accessories for hand tracking, digital gloves and VR interaction.

In late 2023, we noticed a new source of growth in robotics. That prompted me to return my attention to the field and found Noitom Robotics to serve it specifically, working extensively with researchers and robot manufacturers. Looking back, more than a decade of accumulated experience seems to have prepared us for this moment: combining human motion capture with intelligent robot control. It is meaningful work, and I feel fortunate to be doing it.

4. Differences Between Motion-Capture Approaches

Motion capture is the process of digitizing human movement and pose. It is not limited to one technology: computer vision, sensors, magnetic tracking and exoskeletons can all do it. The core is recording and capturing poses and movements, then converting them into digital sequences.

Optical motion capture has long been considered a gold standard in research because it offers absolute spatial accuracy. Here “absolute” refers to calibrated coordinates in the capture space, not zero measurement error. Multiple high-speed industrial cameras use infrared light in the 850–950-nanometer range with global-shutter exposure. All pixels are exposed simultaneously, avoiding the row-by-row timing of a rolling shutter and providing high synchronization accuracy. High-end optical cameras can capture 1,200 frames per second; conventional systems typically operate at 120–240 fps.

For full coverage in a room of roughly fifteen square meters, one or two cameras per square meter are often needed. Professional whole-body capture, as in Hollywood production, might use twenty to thirty cameras in that space. In a larger area of one hundred square meters, even a sparse installation may require forty to fifty cameras, each costing tens of thousands or even more than a hundred thousand yuan. High-quality optical systems can reach submillimeter positional accuracy, and a rigid body formed by several markers can have orientation accuracy of 0.05 degrees. That is why researchers consider it a gold standard for precision.

Optical systems, however, require complex installation and expensive equipment. A complete system often costs four hundred thousand to five hundred thousand yuan. Inertial motion capture offers clear advantages in adaptability and convenience. It needs no external infrastructure and can work indoors or outdoors: put on the equipment, perform a few calibrations and start collecting. The cost is also much lower.

The drawback of inertial capture is its lack of absolute positional accuracy. It estimates motion and pose from wearable sensors measuring angular velocity, acceleration and magnetic fields, and integration introduces cumulative drift. Positional accuracy is generally at the centimeter scale and is relative rather than optical-grade absolute accuracy. Vendors can delay drift through algorithmic improvements, but inertial capture does not provide the same absolute accuracy as optical systems.

In robotics, optical capture is usually chosen when high-precision ground-truth data is needed, such as the absolute position of a person or object. Inertial capture is more common in teleoperation and other scenarios emphasizing flexible environments, ease of use and real-time response.

A more advanced option is hybrid optical–inertial tracking. It combines optical accuracy with inertial continuity, maintaining stable data under occlusion. If one hand blocks the other, the optical system may lose its markers while inertial sensors continue supplying data. Sensor fusion can preserve both accuracy and continuity.

Exoskeletons are another early form of motion capture. In the late 1970s and early 1980s, some research institutions used exoskeleton structures with encoders to track human movement. As robotics research expanded, more researchers returned to exoskeleton-based teleoperation and capture, continuing that earlier idea.

Magnetic motion capture is another option. Some vendors use high-frequency weak magnetic fields and three-axis coils to estimate a sensor's position and orientation, tracking gloves and controllers. Such technology was used by the US Air Force and in other industrial applications, and later appeared in consumer products from companies such as Razer.

Overall, motion capture has evolved for more than half a century, from experimental research systems in the 1970s to film production, sports analysis, virtual reality and today's robotics and embodied intelligence. It has become an important bridge between human movement and the digital world.

5. Choosing a Motion-Capture System for Robotics

From a systematic perspective, we first need to distinguish the purpose of the data: are we pursuing visual plausibility, or task success and manipulation precision?

If the goal is visual plausibility, video-derived data can be useful. Ailing Zeng, a young researcher I know, participated in a very large dataset collecting internet videos; the reference is at the end. Such data can be entirely adequate visually, because projected keypoints are usually accurate from the camera's viewpoint. The problem is missing depth. Monocular video cannot guarantee precision along the depth axis. Only stereo cameras with a sufficiently wide baseline can provide some depth reference, in the account given here. Editorial clarification: monocular methods can estimate depth using motion cues or learned priors; a single RGB view alone does not uniquely specify absolute depth without additional assumptions. The dataset linked in this article is MiraData, a video dataset, rather than a directly measured 3D motion-capture dataset.

If the aim is to produce model outputs or robot behavior that looks plausible—for example, dancing or boxing with a visual resemblance to a human video—motions extracted from monocular footage may meet that need. The goal is for the movement to look similar from the video's viewpoint. That is a comparatively lightweight application.

But highly dynamic movement with physical plausibility demands greater spatial precision: the robot must balance, jump, follow a path that closes back on itself, or interact with its surroundings. Projects such as Unitree's avatar-style teleoperation and Yanjie Ze's TWIST series, linked in the references, describe their platforms as humanoid teleoperation systems for collecting real interaction data. In that setting, spatial precision matters and more specialized motion-capture equipment is needed.

The choice of professional equipment then depends on the environment and the objective, with two broad cases.

In a laboratory or factory, the space is controlled, the budget may be relatively generous and accuracy requirements can be extremely high—for example, collecting coordinated two-handed actions. I recommend optical capture in that case. It is suitable for an indoor installation, accurate and well synchronized, making it ideal for research and high-quality collection. An institution such as Amazon's robotics laboratory building a high-precision experimental system should use optical or hybrid optical–inertial tracking.

In a changing environment, such as outdoors or a temporary setup, rapid deployment, lower cost and flexibility matter more. Inertial capture is then more suitable. Outdoor teleoperation or operation in ordinary, uninstrumented environments is well matched to this approach.

Different settings and needs can lead to completely different optimal choices. Optical, inertial, magnetic and exoskeleton capture each have a reason to exist. That is why all remain in use and continue developing. Choosing a system requires weighing the specific scene and environment case by case.

6. Why Choose Noitom Robotics' Equipment and Data?

First, you do not necessarily have to choose our equipment or data. I am happy to discuss the options and offer fair advice, then let each team decide for itself. The industry is still very young, and we want more people to help solve the data problem together. Data is unquestionably the biggest bottleneck. Language models can collect text across the internet, and autonomous-driving companies can gather large volumes from real vehicles. Robotics has not yet reached that scale.

Real-world data can be divided into several categories. Ultimately, all are needed, and they must work together to provide enough coverage in both volume and quality.

First, real-robot teleoperation data is extremely important. Human-centered or synthetic data cannot fully express the robot's own embodiment. Only physical robot data contains its actual sensing, actuation and configuration characteristics, including manufacturing differences between models. It is expensive, inflexible and difficult to transfer across embodiments, but I believe every robot model still needs at least several thousand hours of real teleoperation data. This is like the automotive industry collecting substantial new road-test data for each vehicle model. A robot's configuration or hardware changes also require new collection.

Real-robot teleoperation is therefore indispensable, but cost and efficiency constrain its scale. I see its main role as aligning a model with embodiment-specific characteristics, rather than supplying most of the skill-learning data. Skills and regularities should primarily come from other sources.

Researchers and companies have increasingly turned to human-centered collection. Directly recording human interaction and visual information provides rich examples for robot imitation and understanding. This includes the largely visual interaction data represented by projects such as Ego4D, EgoDex, EgoPressure and Project Aria or Aria 2.

There is disagreement over what human-centered data means. Some researchers emphasize in-the-wild collection to ensure freedom and diversity. But genuinely natural collection requires sacrificing some modalities, dimensions or precision. A complex, burdensome wearable system cannot readily enter homes, factories or care homes. Some teams try to have everything—huge scale, rich modalities and high precision—and encounter contradictory requirements.

Human-centered data may therefore need its own layers. The upper layer is comparatively smaller, highly precise, multimodal data, potentially on the scale of one hundred thousand to one million hours, establishing a high-quality benchmark across modalities and dimensions. Below it is a larger volume of in-the-wild data, with simpler wearables and lower precision requirements, aiming for broader scene coverage.

Below that is another crucial layer: synthetic and internet data. It can augment real-robot and human-centered data at large scale. Simulation and synthetic environments expand the diversity of trajectories, goals, backgrounds and strategies, multiplying the data volume by hundreds of times. Autonomous driving used this approach early on, extending real data through simulation to cover extreme situations.

I do not start from the question of how to promote our equipment or datasets. I want the industry to work together on the data bottleneck. Without a complete data system built collaboratively, progress will stall. Larger-scale, more advanced data-driven approaches are needed to move robot intelligence forward.

7. Near-Term Contributions to the Community

Noitom Robotics is not simply an equipment manufacturer, and we do not want to be only that. Noitom's earlier business can in many respects be seen as equipment manufacturing, but Noitom Robotics is different. We want to address the industry's data problem in multiple ways, acting as a service provider and a dataset licensor as well.

For a considerable time, I will use equipment to make friends rather than profit from selling it. I will provide it free to many researchers. For strong industry partners, we may offer it at cost or partly as a gift. The purpose is to let more people use the technology and advance research and applications.

I will devote substantial effort and capital to building our own data factory, accumulating datasets of one hundred thousand or even one million hours. The aim is to lower costs across the industry. That may sound contradictory: why would a third-party data company make collection cheaper? My logic is that only large-scale data that can transfer across embodiments and be reused is worth accumulating.

We will find effective ways to align our human-centered datasets with the different robot configurations used by partners and customers, then license the data to them. Cross-embodiment mapping allows repeated licensing and reuse, substantially reducing duplicated investment and waste across the industry.

For example, we recently reached a data-licensing agreement with a large company on the scale of tens of millions in value. The licensed price was even slightly below the cost of collection. But I also spent a long time persuading the company to give up complete exclusivity. That business model can make industry-wide data accumulation more efficient and sustainable.

I believe this is the greatest value we can provide. We need reasonable revenue to sustain an R&D team of dozens to more than a hundred people and keep investing in technology. At the same time, reusable datasets let the whole industry obtain high-quality resources at far less than the cost of collecting them independently.

That is my vision, and it will take time to put it into practice. I will focus on this over the next two or three years. Robot manufacturers, model companies and investors interested in the direction are welcome to discuss it with me. I am confident we can find suitable ways to work together.

8. Talent Needs

Our greatest need at present is people. We are urgently recruiting in three categories.

First are product managers and project managers. We need people in both data and robotics, whether junior or experienced. Anyone committed to this direction is welcome.

Second are algorithm specialists. Our chief scientist, Lei Han, is recruiting researchers with experience in multimodal models, VLA, reinforcement learning and data science. We have openings for young researchers, interns and full-time research scientists and hope to attract people who truly care about robotics algorithms and models.

Third are sensor and hardware specialists. We have accumulated hardware experience, but that is only a starting point. Most motion-capture and tracking equipment on the market was designed for film effects or games, not robot data collection. Yet the robot-data market is much larger than the film market and has higher equipment requirements. It needs new hardware designed for different environments: in the wild, in factories and in research labs.

We are actively pursuing that direction and especially need hardware-planning and sensor expertise: magnetic-positioning specialists, IMU specialists, optical-sensor experts, engineers who turn algorithms into products, firmware engineers and BSP, or board-support-package, engineers. If you have experience in these areas or want to find an opportunity in a promising field, we welcome you.

9. Business Partnerships

Let me also take this opportunity for a straightforward commercial pitch. If your company or institution does not need to do everything internally, we would be pleased to partner with you. Companies such as Tesla and Figure pursue full vertical integration, sometimes even saying they want to build their own VR systems. That is admirable, but it does not suit every company.

If you have a clear boundary around your core business and want a specialist team to handle important supporting work such as data collection or dataset preparation, we would welcome the collaboration. With the most reasonable pricing, the greatest enthusiasm and the highest professional standards, we can help you complete the work efficiently.

References

The TWIST Series

https://arxiv.org/abs/2505.02833

https://arxiv.org/html/2511.02832

MiraData, a Dataset Coauthored by Ailing Zeng

https://arxiv.org/abs/2407.06358

Ego4D: https://arxiv.org/abs/2110.07058

EgoDex: https://arxiv.org/abs/2505.11709

EgoPressure: https://arxiv.org/abs/2409.02224

Smart Note · Human Data