What kind of data does a robot need to learn an operation? First-person video can record the objects and environment in front of it, but it struggles to fully reflect how the body moves or how the hand contacts an object.
According to Sina Tech, Poxiao Intelligence released SeePhys Hub on October 14, combining a headband, wrist devices, waist and leg trackers, and tactile gloves, together with a training data processing platform, aimed at the data collection needs of robot learning.
From “Seeing the Process” to Recording Motion and Contact
Take picking up an object as an example: video can show the process of the hand reaching toward the target and the object leaving the tabletop. But to understand the operation, it may also be necessary to know how the body posture adjusts, how the wrist rotates, and where contact occurs and how the force changes.
This also explains why full-body motion and tactile data are drawing attention: they provide information beyond vision, helping to describe an operation more completely. The example above is an explanation of what the data can be used for, not a verified specific task result that SeePhys Hub has already achieved.
After Collection, the Data Must Also Be Made Usable for Training
Multiple devices recording simultaneously does not mean the data can be put directly into training. Different signals need to correspond in time, device positions need to be calibrated, and missing or anomalous records need to be identified and handled. Otherwise, the contact actions in the footage and the sensor records may not match accurately.
Therefore, evaluating this type of system involves not only wearing convenience and continuous collection capability, but also data synchronization, the processing workflow, and the proportion of data that is ultimately usable. The relevant performance should still be understood as company-disclosed information for now, and a product launch cannot be equated with having achieved large-scale delivery.
This launch of SeePhys Hub places the focus of observation on the upstream link of robot learning: how to record human operating experience. How much it can help robot training still needs to be judged in light of data quality, training methods, and specific task results.
Source: Sina Tech