With hand tracking they mean finger tracking, seeing you 'click' with your index finger and thumb. The Oculus has hand tracking today via controllers already.
They mean handtracking. Oculus touch is a tracked controller, which means the controller is tracked and the tracking system has no concept of a hand. You could put the controller on your foot, or the floor or on your dog and it would track in the same way as your hand. Hand tracking uses computer vision to recognize hands, for most humans fingers are included in that
And Quest has that proper hand tracking already (had it for something like 5 months now), it's just been an experimental feature until now. I haven't used it much; for some reason the tracked hands seem to be placed lower than the fake hands when I'm using controllers, which makes the latter a better experience for me.
Increased latency due to using complex, deep learning based, entirely computer vision driven hand tracking vs. tracked controllers which partially use accelerometer data to correct drift and account for latency, and use a much faster classic computer vision algorithm for position tracking (https://developer.oculus.com/blog/tracking-technology-explai...)
Do Deep Learning-based methods introduce all that much latency? My understanding of neural network is that the training is expensive, but using them is essentially free - you have to push your pixels through a few layers of matrix multiplications and you have the results; it's just tuning those matrices that takes time.
VR latency is measured in milliseconds, the operation you described can easily take multiple frames to complete even with a very fast, optimized model. Machine learning inference is never free, and often isn't real-time. Training can take hours to weeks depending on the model.
