PHANES AI has released SeePhys Hub, a whole-body, multimodal data capture system meant to turn everyday human activity into training data for robots. QbitAI reported that the October 14 launch was a joint release with Tinno Mobile, which contributed experience in cost control, supply chain and quality assurance, while PHANES AI focused on product design and algorithms.
The kit consists of a four-camera headband, two wrist bands, three motion trackers worn on the waist and lower legs, and a pair of tactile gloves. Together they record first-person vision with 360-degree coverage, hand pose and high-density touch, and whole-body pose and force.
The company's argument is that egocentric, or Ego, data from head-mounted cameras has become a key source for robot learning because it scales easily, but it misses much of what the body is doing. A camera at eye level sees the hands, yet often not the elbows or legs, and it records nothing about how hard the hands press on an object. Tasks that require walking, bending and two-handed coordination need that missing information.
SeePhys Hub uses a sparse-node approach. The three trackers combine IMU motion signals with UWB ranging between nodes, and algorithms fuse them with head and wrist vision to estimate full-body pose. The headband weighs no more than 220 grams and each wrist band no more than 75 grams; both can record continuously for 24 hours. Sync error among the headband's cameras, IMU and UWB is under 0.5 milliseconds.
The gloves are optional. Founder and CEO Yang Shuo said data collection operators asked the team not to make their workers wear gloves, since they can hinder delicate tasks such as dishwashing and can introduce sensor noise. Customers can choose configurations by task, and the team is researching how to estimate touch from vision.
A companion platform handles parsing, cleaning, segmentation and semantic labeling, and links to scene reconstruction, pose estimation, hand tracking and motion retargeting onto robot bodies. Yang, who became a tenured professor at Harbin Institute of Technology (Shenzhen) at 26, described the device as aimed at a middle layer of data, between internet video and high-precision motion capture or real-robot data.
SEE ALSO: MagicLab's EchoPolicy-0 Tops AI2's MolmoSpaces Combined Board With a 75.4% Manipulation Success Rate