Dawn Intelligence and Tinno Mobile Launch SeePhys Hub Full-Body Multimodal Data Collection System

Over the past few years, robot data collection has been doing subtraction: moving from teleoperation toward Ego, shedding the robot body and simplifying collection equipment, so that more everyday human operations can become training data obtainable at scale. But Dawn Intelligence believes that when robots begin learning to step, bend, carry, and perform fine manipulation, the body motion and contact information beyond the camera's view becomes a part that cannot be easily discarded.

Vision-dominant Ego data can see what a person completed, but it cannot necessarily fully record the body state behind the action. When a person raises a hand to reach an item on a high shelf, the shoulder, arm, and torso need to adjust together; a first-person camera can capture the hands, but it cannot necessarily see the position of the elbows and legs; when encountering self-occlusion, object occlusion, or view blind spots, the reliability of full-body pose estimation becomes a problem.

SeePhys Hub's positioning is precisely to fill this gap. On October 14, Dawn Intelligence partnered with Tinno Mobile to release this full-body, full-modality data collection system. The entire setup consists of a four-camera head ring, two wrist bands, three waist-and-leg motion trackers, and a tactile glove, recovering full-body pose through fusion of vision and motion signals, and collecting hand contact pressure according to task requirements.

SeePhys Hub did not take the route of stacking sensors, but instead adopts a sparse-node approach: placing three motion trackers at the waist and both lower legs, combined with head vision and dual-wrist information, recovering full-body pose through algorithms. The IMU provides motion-related signals, UWB provides ranging information between nodes, and these are then combined with vision to estimate full-body pose.

Visual coverage has also been expanded. The four-camera head ring provides 360° surround vision, while the wrist cameras supplement close-range manipulation details. The R&D team gave an example: after taking a bottle from the refrigerator, a person turns sideways to close the door, and the camera behind the head can still record this action—operations in real scenarios often happen amid exactly this kind of turning, moving, and view change.

Portability was placed in an important position in product design: the head ring weighs no more than 220 grams, and a single wrist band no more than 75 grams. The team regards portability, cost, and freedom of activity in real scenarios as the prerequisites for whether collection at scale can hold up.

Multiple signal channels also need to be aligned to the same moment. The instant a hand touches an object in the video needs to correspond to pressure changes and motion states, so that the complete interaction process can later be analyzed. According to product materials, the synchronization precision of the cameras, IMU, and UWB inside the head ring is no more than 0.5 milliseconds. This inconspicuous engineering step directly affects whether data from different modalities can be effectively connected.

After collection there is also a processing step. The companion platform for SeePhys Hub handles parsing, cleaning, motion segmentation, and semantic annotation, and connects to scene reconstruction, pose estimation, and hand tracking; motion retargeting then maps human motion onto the robot body, providing support for subsequent training and adaptation.

It should be noted that parameters such as the 0.5-millisecond synchronization precision, the 220-gram head ring, and the 75-gram wrist band all come from product materials or the team's own statements, and the report gives no third-party verification or quantitative results. These numbers can illustrate design goals, but they are not yet sufficient to prove actual collection performance.

The tactile glove records pressure changes when both hands contact objects, giving the actions in the video a corresponding physical signal. But it is an optional design, and the bands and glove can be used separately, with customers choosing configurations according to the task.

This trade-off comes from real feedback. According to Yang Shuo, the team has been in contact with some data collection operators, and "the first requirement they gave us was not to make our people wear gloves." The reason is that in specific scenarios involving water resistance, such as washing dishes, fine manipulation is affected by what is worn, and the sensors themselves may also produce noise.

The team once encountered in testing that after the glove had grabbed a water cup several times, an open bare hand still showed localized pressure. When this kind of noise enters the data, it increases the difficulty of subsequent processing and training. If wearable equipment affects a person's operation, it also affects the value of the record. The team says it is still researching how to estimate touch from vision, in order to expand the usable sources of data.

In the three-tier data pyramid described by Yang Shuo, the bottom layer is the large volume of internet video, with pose and tactile information supplemented through algorithms; the top layer is data such as high-precision motion capture and real-machine operation; SeePhys Hub targets the middle layer, attempting to balance the potential for collection at scale with a more complete record of interaction.

Each layer takes on a different task: motion capture data can be used to supervise pose estimation, real-machine data helps the model adapt to robots, and human video provides more scenarios and operations. What wearable devices supplement is the ability to go into real environments and continuously record alongside a person.

To truly roll out the devices, supply chain and quality assurance are also part of the picture. In this collaboration, Dawn focuses on product design and algorithms, while Tinno Mobile provides experience in cost control, supply chain, and quality assurance.

According to the team's plan, customers with algorithmic capabilities can buy the hardware and process data themselves; customers needing a complete solution can choose the companion algorithms and local deployment; scenarios such as research can also use the cloud platform to obtain processed data. Data collection operators, embodied model teams, scenario parties, and research institutions are all users targeted by the SeePhys Hub team.

In Yang Shuo's view, the accumulation behind the device also includes pose estimation algorithms, as well as paired data used to train and validate algorithms. Only by making hardware, algorithms, and data processing work together can the sensors worn on a person's body be turned into operation records usable for training.

These commercial arrangements currently all come from the team's plans, and no actual deployment scale or customer validation results have been disclosed.

"Our ultimate goal is definitely to build the model well." Yang Shuo gave this answer when explaining why Dawn is developing data collection devices. In the team's view, the problem lies in what data the model needs: to model body motion, hand contact, and environmental change together, corresponding operation records are needed, and the required data is difficult to obtain at low cost and at scale, so the team chose to fill in this entry point itself.

According to the plan, Dawn Intelligence expects to launch a full-modality world model in 2027, and to prioritize application validation in vertical scenarios such as industry and food service retail. Whether this timetable can be realized on schedule is not given further basis in the report, and uncertainty remains.

At this stage Dawn still does not plan to build its own robot body. Yang Shuo's consideration is that data, models, and deployment scenarios are still changing, and whether the robot needs to carry heavy objects and how much payload it needs both affect body design; settling on hardware too early could also limit subsequent algorithm directions, a risk he calls "pure sunk cost." The team currently uses general-purpose bodies to test models and plans to cooperate with body makers and scenario parties on application pilots; whether it will launch its own body in the future depends on scenario needs and commercial returns becoming clearer.