Haptics can greatly degrade with lag. There have been some approaches to predict what the operator is going to do or predict that a collision is going to occur and prevent such things, but they have not been able to decrease the time it takes to complete a task[0].
It is worth noting though that remote surgery has been somewhat successful over high bandwidth communication links. One constraining factor for teleoperation research is that there aren't many applications where it makes sense to do so. The problem is that robots and haptic gear are currently expensive which limits the applications. You need an application where humans cannot go or the labor the operator is doing is more valuable than the cost of transporting the operator there and the robot. Applications where humans can't go are fairly niche. In addition, applications like working inside a non-damaged nuclear reactor or subsea operations tend to involve equipment that has been designed to be serviced by the simplest of teleoperated robots. In the case of nuclear reactors, mechanical and hydraulic remote manipulators are suitable for most operations. Very valuable tasks like robot surgery are rarely done remotely, partially because of how expensive surgical robots are. We could put surgical robots in remote locations, but remote locations have less patients to service, so it's hard to justify the cost of putting robots in a bunch of remote locations versus moving the patients to the hospitals.
[0]https://ieeexplore.ieee.org/document/7399376/ [1]https://en.wikipedia.org/wiki/Remote_surgery
JAXA has been looking into this. It seems that highly motivated operators can tolerate a full second of lag.
In the DARPA robotics challenge, competitors were able to create movable views with lag by 3d scanning the environment and streaming the 3d scans.
That's entirely suitable for a "work bench." In fact, the background might as well be blanked out and replaced with a synthetic grid.
You need an application where humans cannot go or the labor the operator is doing is more valuable than the cost of transporting the operator there and the robot.
I can envision many contexts in the upcoming decades where teleoperation is going to be very useful. The long term health costs of human beings suiting up to go into environments without protection from radiation could be quite high.
Why isn't teleoperation a more heavily researched area? I suspect that it could be efficiently accomplished with manipulators that are much simpler than human beings or human hands, so long as you add two powerful ingredients:
1) Movable point of view. Basically this could be slaved to the position of a high resolution VR headset. It could even be provided by a tethered drone.
We've had teleoperation for a long time, but it has generally been hampered by a static point of view. What if you could crane your neck and look around what you are doing? This greatly improves the intuitiveness and effective bandwidth of the interface.
2) Haptic feedback.
If you can feel what's happening, even indirectly, it also greatly improves the quality of the interface. Think of the dexterity you have with a pair of forceps or needle nose pliers, versus the dexterity you'd have with those while your hands are numbed by novacaine. Since such mechanisms would be much simpler than a mimic human arm, they'd be much cheaper. Also, human beings are very adaptable to control tasks and don't really require a precise mimic of human limbs. (Witness the amazing feats of backhoe control you can find on YouTube.)