PHANES AI Launches SeePhys Hub to Capture Whole-Body Motion and Touch for Ego Data

On October 14, PHANES AI and Tianlong Mobile released SeePhys Hub, a whole-body multimodal data capture system. The kit pairs a four-camera headband with two wrist bands, three waist and calf motion trackers and an optional haptic glove, fusing visual and motion signals to recover full-body pose while recording hand contact pressure when a task calls for it.

Egocentric, or Ego, data has become a major source for embodied AI: a person wears a device, goes about ordinary tasks and a camera follows them, so collection no longer has to happen around a robot body. But for tasks that involve moving and coordinating the whole body, recording only what is in front of the wearer leaves out key state. Reaching for something high requires shoulders, arms and torso to move together, and a head-mounted camera that captures both hands may never see the elbows or the legs.

SeePhys Hub's answer is to add a small number of tracking nodes at key points on the body. Three trackers sit at the waist and both calves, and their data is combined with head-mounted vision and wrist information to reconstruct full-body pose. IMUs supply motion signals, UWB provides ranging between nodes, and everything is fused with vision. The headband offers 360-degree surrounding vision, wrist cameras supply close-up manipulation detail, and the optional haptic glove records pressure where the hands meet objects.

Aligning the streams matters as much as collecting them. According to product materials, the cameras, IMU and UWB inside the headband are synchronized to within 0.5 milliseconds, so the instant a hand touches an object lines up with the pressure change and motion state, which is what makes a full interaction analyzable later. A companion platform then handles parsing, cleaning, action segmentation and semantic annotation, and connects to scene reconstruction, pose estimation, hand tracking and motion retargeting.

The engineering trade-offs lean toward wearability: the headband weighs no more than 220 grams and each wrist band no more than 75 grams. Making the haptic glove optional follows the same logic. Founder and CEO Yang Shuo says some data-collection operators asked first that their staff not have to wear gloves at all, because waterproof tasks such as washing dishes are harder to do with them on and the sensors themselves can produce noise; in testing, the glove sometimes reported residual pressure after a few runs of holding a cup.

In Yang Shuo's three-layer data pyramid, internet video sits at the bottom with pose and touch estimated by algorithms, high-precision motion capture and real-robot data sit at the top, and SeePhys Hub targets the middle — trying to combine the scalability of light capture with a fuller record of interaction. Under the partnership, PHANES AI handles product design and algorithms while Tianlong Mobile contributes cost control, supply chain and quality assurance experience.

The team's earlier work traces a similar path through touch: it released the EgoTouch dataset and the TouchAnything study in May 2026, recording first-person video, hand pose and real contact pressure together and then estimating contact and pressure distribution from video. TouchWorld later let touch participate in robot action, with models predicting visual and contact targets and correcting deviations using real haptic feedback. SeePhys Hub extends the scope from hand contact to whole-body motion, which is what the team argues whole-body mobile dexterous manipulation requires.

Under its roadmap, PHANES AI expects to release a whole-body multimodal world model in 2027 and validate it first in verticals such as industry and food service and retail; it is not building its own robot body yet, with Yang Shuo calling a premature hardware bet pure sunk cost. For teams training whole-body mobile manipulation, whether the data funnel can scale may matter more to progress than model parameter counts.