The Goal – Explain What is Magic Leap Doing
Magic Leap has a way of talking about what they hope to do someday and not necessarily what they can do anytime soon. Their patent applications are full of things that are totally impossible or impractical to implement. I’ve been reading well over a thousand pages of Magic Leap (ML) patents/applications, various articles about the company, watching ML’s “through the optics” videos frame by frame, and then applying my own knowledge of display devices and the technology business to develop a picture of what Magic Leap might produce.
Some warnings in advance
If you want all happiness and butterflies, as well as elephants in your hand and whales jumping in auditoriums, or some tall tale of 50 megapixel displays and of how great it will be someday, you have come to the wrong place. I’m putting the puzzle together based on the evidence and filling in with what is likely to be possible in both the next few years and for the next decade.
Separating Fact From Fiction
There have been other well meaning evaluations such as “Demystifying Magic Leap: What Is It and How Does It Work?“, “GPU of the Brain“, and the videos by “Vance Vids” but these tend to start from the point of believing the promotion/marketing surrounding ML and finding support in the patent applications rather than critically evaluating them. Wired Magazine has a series of articles as well as Forbes and others have covered ML, but these have been are personality and business pieces that make no attempt to seriously understand or evaluate the technology.
Among the biggest fantasies surrounding Magic Leap is the Arrayed Fiber Scanning Displays (FSD); many people think this is real. ML Co-founder and Chief Scientist, Brian Schowengerdt, develop this display concept at the University of Washington based off an innovative endoscope technology and it features prominently in a number of ML assigned patent applications. There are giant issues in scaling up FSD technology to high resolution and what it would require.
In order to get on with what ML is most likely doing, I have moved to the Appendix why FSDs, light fields, and very complex waveguides are not what Magic Leap is doing. Once you get rid of all the “noise” of the impossible things in the ML patents, you are left with a much better picture of what they are actually could be doing.
What left is enough to make impressive demos and it may be possible to produce at a price that at least some people could afford in the next two years. But ML still has to live by what is possible to manufacture.
Magic Leaps Optical “Magic” – Focus Planes
At the heart all of ML optical related patents is the concept eye vergence-accomodation where the focus of the of the various parts of a 3-D image should agree with their distances or it will cause eye/brain discomfort. For more details about this subject see this information about Stanford’s work in this area and their approach of using quantized (only 2 level) time sequential light fields.
There are some key similarities in that between the Stanford and Magic Leap’s approaches. They both quantize to a few levels to make them possible to implement and they both present their images time sequentially and they rely on the eye/brain to both fill in between the quantizated levels and integrate a series of time sequential images. Stanford’s approach is decidedly not a “see through” with an Oculus-like setup with two LCD flat panel displays in series where Magic Leap’s goal is to merge the 3-D images with the real world with Mixed Reality (MR).
Magic Leap uses the concept of “focus planes” where they conceptually break up a 3-D image into quantized focus planes based on the distance of the virtual image. While they show 6 virtual planes in Fig. 4 from the ML application above, that is probably what they would like to do but they are doing fewer planes (2 to 4) due to practical concerns.
Magic Leap then renders the parts of an image image into the various planes based on the virtual distance. The ML optics make it planes appear to the eye like they are focus based their corresponding virtual distance. These planes are optically stacked on top of each other give the final image and they rely on the person’s eye/brain to fill in for the quantization.
Frame Sequential Focus Planes With SLMs
Magic Leap’s patents/applications show various ways to generate these focus planes, the most fully form concepts use a single display per eye and present the focus planes time sequentially in rapid succession, what ML refers to as “frame-sequential“where there is one focus plane per “frame.”
Both due to the cost and size multiple displays per eye and their associated optics including those to align and overlay them, the only possible way ML could build a product for even a modest volume market is by using frame sequential methods using a a high speed spatial light modulator (SLM) such a DLP, LCOS, or OLED microdisplay.
Waveguides and Focus Planes
Light rays that coming from a far away point that make into the eye are essentially parallel (collimated) and light rays from a near point have a wider set angles. These differences in angles is what makes them focus differently, but at the same time creates problems for existing waveguide optics, such as what Hololens is using.
The very flat and thin optical structures call “waveguides” will only work with collimated light entering them because of how total light totally internally reflects to stay in the light guide and the the way the diffraction works to make the light exits. So a simple waveguide would not work for ML.
Some of ML’s concepts use use one or more beam splitting mirrors type optics rather than waveguides for this reasons. Various ML’s patent applications show using a single large beam splitter or multiple smaller ones (such as at left), but these will be substantially thicker than a typical waveguide.
What Magic Leap calls a “Photonics Chip” looks to be at least one layer of diffractive waveguide. There is no evidence of mirror structures, and because it bends the wood in the background (if it were just a simple plate of glass, the wood in the background would not be bent), it appears to be a diffractive optical structure.
Because ML is doing focus planes, they need to have not one, but a stack of waveguides, one per focus plane. The waveguides in ML’s patent applications show collimated light entering the each waveguide in the stack like a normal waveguide, but then the exit diffraction gratings both causes the light to exit also imparts the appropriate focus plane angle to the light.
To be complete, Magic Leap has shown in several patent applications shown some very thick “freeform optics” concepts, but none of this would look anything like the “Photonics Chip” that ML shows. ML’s patent applications show many different optical configurations and they have demoed a variety of different designs. What we don’t know is if the Photonics Chip they are showing is what they hope to use in the future or if this will be in their first products.
Magic Leaps Fully Formed Designs In Their Recent Patent Applications
Most of Magic Leaps patent applications showing optics have more like fragments of ideas. There are lots of loose ends and incomplete concepts.
More recently (one publish just last week) there are patent applications assigned to Magic Leap with more “fully formed designs” that look much more like they actually tried to design and/or build them. Interestingly, these applications don’t include as inventors the founders Rony Abovitz, the CEO, nor even Brian T. Schowengerdt, Chief Scientist, while they may use ideas from those prior “founders patent application.”
While the earlier ML applications mention Spatial Light Modulators (SLMs) using DLP, LCOS, and OLED microdisplays and talk about Variable Focus Element (VFEs) for time sequentially generating focus planes, they don’t really show how to put them together to make anything (a lot is left to the reader).
Patent Applications 2016/0011419 (left) and 2015/0346495 (below) show straight forward ways to achieve field sequential focus planes using a Spatial Light Modulator (SLM) such as DLP, LCOS or OLED microdisplay.
As focus plane is created by setting the a variable focus element (VFE) to a one focus point and then generating the image by the SLM. Then the VFE focus is then changed and a second focus plane is displayed by the SLM. This process can be repeated to generate more focus planes and limited by how fast the SLM can generate image and by level of motion artifact that can be tolerated.
These are clearly among the simplest way to generate focus planes. All that is added over a “conventional” design is the VFE. When I first heard about Magic Leap many months ago, I heard they were using DLPs with multiple focus depths but a more recent Business Insider is reporting ML is using using Himax LCOS. Both of these could easily be adapted to support OLED microdisplays.
The big issue I have with the straight forward optical approaches are the optical artifacts I have seen in the videos and the big deal ML makes out of their Photonics Chip (waveguide). Certainly their first generation might use a more straightforward optical design and then save the Photonics Chip for the next generation.
Magic Leaps Videos Show Evidence of Waveguide Optics
As I wrote last time, there is a lot of evidence from the videos ML has put out that they are using a waveguide at least for the video demos. The problem is when you bend light in a short distance using diffraction gratings or holograms is that some of the light does not get bent correctly and this shows up colors not lining up (chroma aberrations) as well as what I have come to call the “waveguide glow”. If at R2D2 below (you may have to click on the image see it clearly) you should see a blue/white glow around R2D2. I have seen this kind of glow in every diffractive and holographic waveguide I have seen. I have heard that the glow might be eliminated someday with laser/very narrow bandwidth colors and holographic optics.
The point here is that there is a lot of artifact evident that ML was at least using some kind of waveguide in their videos. This makes it more likely that their final product will also use waveguides and at the same time may have some or all of the same artifacts.
Best Fit Magic Leap Application with Waveguides
If you drew a venn diagram of all existing information, the one patent application that fits best it all is the very recent US 2016/0327789. This is no guarantee that it is what they are doing, but it fits the current evidence best. It combines the a focus plane sequential LCOS SLM (although it shows it could also support DLP but not OLED) with waveguide optics.
The way this works is that for every focus plane there are 3 Waveguides (RED, Green,and Blue) and spatial separate set of LEDs Because the are spatially separate, they will illuminate the LCOS device at a different angle and after going through the beam splitter the waveguide “injection optics” will cause the light from the different spatially separated LEDs to be aimed at a different waveguide of the same color. Not shown in the figure below is that there is an exit grating that both causes the light to exit the waveguide and imparts an angle to the light based on the focus associated with that give focus plane. I have coloring in the “a” and “b” spatially separated red paths below (there are similar pairs for blue and green).
With this optical configuration, the LCOS SLM is driven with the image date for a given color for a given focus plane and then the associated color LED for that plane is illuminated. This process then continues with a different color and/or focus plane until all 6 waveguides for the 3 colors by 2 planes have been illuminated.
The obvious drawbacks with this approach:
- There are a lot of layers of waveguide with exit diffraction gratings that the user will be looking through and the number of layers grows by 3 with each added focus plane. That is a lot of stuff to be looking though and it is bound to degrade the forward view.
- There are a lot of optical devices that all the light is passing through and even small errors and leak light builds up. This can’t be good for the overall optical quality. These errors have their effect on resolution/blurring, chroma aberrations, and glowing/halo effects.
- Being able to switch though all the colors and focus planes fast enough to avoid motion artifacts where the colors and/or the focus planes break up. Note this issue exist with using any approach that both does field and focus plan sequential. Obviously this issue becomes worse with more focus planes.
The ‘789 patent show an alternative implementation for using a DLP SLM. Interestingly, this arrangement would not work for OLED Microdisplays as they generate their own illumination so you would not be able to get the spatially separated illumination.
So what are they doing?
Magic Leap is almost certainly using some form of spatial light modulator with field sequential focus planes (I know I will get push-back form the ML fans that want to believe in the FSD — see the Appendix below); but this is the only way I could see them going to production in the next few years. Based on the Business Insider information, it could very well be an LCOS device in the production unit.
The the 2015/0346495 with the simple beam splitter would be what I would have choose for a first design provide there is an appropriate variable focus element (VFE) available. It is by far the simplest design and would seem to have the lowest risk. The downside is that the angled large beamsplitter will make it thicker but I doubt that much more so. Not only is it lower risk (if the VFE works) but the image quality will likely be better using a simple beam splitter and spherical mirror-combiner than many layers diffractive waveguide.
The 2016/0327789 application touches all the basis based on available information. The downside is that they need 3 waveguides per focus plane. So if they are going to say support just 3 focus planes (say infinity, medium, and short focus) they are going to have 9 (3×3) layers waveguides to manufacture and pay for and 9 layers to look through to see the real world. Even if each layer is extremely good quality, the error will build up in so many layers of optics. I have heard that the Waveguide in Hololens has been a major yield/cost item and what ML would have to build would seem to be much more complex.
While Magic Leap certainly could have something totally different, but they can’t be pushing on all fronts at once. They pretty much have to go with a working SLM technology and get their focus planes time sequentially to build an affordable product.
I’m fond to repeating the 90/90 rule that “it takes 90% of the effort to get 90% of the way there, then it takes the other 90% to do the last 10%” and someone quipped back, it can also be 90/90/90. The point being is that you can have something that look pretty good and impresses people, but solving the niggling problems, making it manufacturable and cost effective almost always takes more time, effort, and money than people want to think. These problems tend to become multiplicative if you take on too many challenges at the same time.
Comments on Display Technologies
As far as display technologies go each of the spatial light technologies has it pro’s and cons.
- LCOS seems to be finding the widest acceptance due to cost. It is generally lower power in near eye displays than DLP. The downside is that it has a more modest field rate which could limit the number of focus planes. It could also be used in any of the 3 prime candidate optical system. Because the LEDs are separate from the display, they can support essentially any level of brightness.
- DLP has the fastest potential field rate which will support more focus planes. With DLPs they could trade color depth for focus planes. DLPs will also tend to have higher contrast. Like LCOS, brightness will not an issue as the LEDs can provide more than enough light. DLP tends to be higher in cost and power and due to the off axis illumination, tend to have a little bigger optical system that LCOS in near eye applications.
- OLED – It has a lot of advantages in that it does not have to sequentially change the color fields, but the current devices still have a slower frame rate than DLP and LCOS can support. What I don’t know, is how much the field rate is limited by the OLED designs to date versus what they could support if pressed. The lack of control of the angle of illumination such as used in the ‘789 application. OLEDs put out rather diffuse with little angle control and this could limit its usefulness with respect to focus plane where you need to control the angles of light.
- FSD Per my other comment and the Appendix below, don’t hold your breath waiting for FSDs.
Image Quality Concerns
I would be very concerned about Magic Leap’s image quality and resolution beyond gaming applications. Forget all those magazine writers and bloggers getting all geeked out over a demo with a new toy, at some point reality must set in.
Looking at what Magic Leap is doing and what I have seen in the videos about the effective resolution and image quality it is going to be low compared to what you get even on a larger cell phone. They are taking a display device that could produce a good image (either 720p or maybe 1080p) under normal/simple optics and putting it through a torture test of optical waveguides and whatever optics used to generate their focus planes at a rational cost; something has to give.
I fully expect to see a significant resolution loss no matter what they do plus chroma aberrations, and waveguide halos provide they use waveguides. Another big issue for me will be the “real world view” through whatever it takes to create the focus planes and how will it effect you say seeing you TV or computer monitor through the combiner/waveguide optics.
I would also be concerned about field sequential artifacts and focus plane sequential artifacts. Perhaps these are why there are so many double images in the videos.
Not to be all doom and gloom. Based on casual comments from people that have seen it and the fact that some really smart people invested in Magic Leap, it must provide an interesting experience and image quality is not everything for many applications. It certainly could be fun to play with at least for a while. After all, Oculus rift has a big following and its angular resolution is so bad that they cover up by blurring and it has optical problems like “god rays.”
I’m more trying to level out the expectations. I expect it to be a long way from replacing your computer monitor, as one reporter suggested, or even your cell phone, at least for a very long time. Remember that this has so much stuff in that in addition to the head worn optics and display you are going to have a cable down to the processor and battery pack (a subject I have only barely touched on above).
Yes, Yes, I know Magic Leap has a lot of smart people and a lot of money (and you could say the same for Hololens), but sometime the problem is bigger than all the smart people and money can solve.
The Big Things Magic Leap is NOT Going To Make in Production Anytime Soon
The first step in understand Magic Leap is to remove all the clutter/noise that ML has generated. As my father use to often say, there are to ways to hide information, you can remove it from view or your can bury it.” Below is a list of the big things that are discussed by ML themselves and/or in their patents that are either infeasible or impossible any time soon.
It would take a long article on each of these to give all the reasons why they are not happening, but hopefully the comments below will at least outline the why:
A) Laser Fiber Scanning Display (FSD)
A number of people of picked up on this particularly because the co-founder and Chief Scientist, Brian Schowengerdt, developed this at the University of Washington. The FSD comes in two “flavors” the low resolution single FSD and the Arrayed FSD
1) First, you pretty limited on the resolution of a single mechanically scanning fiber (even more so than Mirror scanners). You can only make them spiral so fast and they have their own inherent resonance. They make an imperfectly space circular spiral that you then have to map a rectangular grid of pixels onto. You can only move the fiber so fast and you can trade frame rate for resolution a bit but you can’t just make the fiber move faster with good control and scale up the resolution. So maybe you get 600 spirals but it only yields maybe 300 x 300 effective pixels in a square.
2) When you array them you then have to overlap the spirals quite a bit. According to ML patent US 9,389,424 it will take about 72 fibers scanner to made a 2560×2048 array (about 284×284 effective pixels per fiber scanner) at 72 Hz.
3) Lets say we only want 1920×1080 which is where the better microdisplays are today or about 1/2.5 of 72 fiber scanners or about 28 of them. This means we need 28 x 3 (Red, Green, Blue) = 84 lasers. A near eye display typical outputs between 0.2 and 1 lumen of light and you divide this then by 28. So you need a very large number really tiny lasers that nobody I know of makes (or may even know how to make). You have to have individual very fast switching lasers so you can control them totally independently and at very high speed (on-off in the time of a “spiral pixel”).
4) So now you need to convince somebody to spend hundreds of millions of dollars in R&D to develop very small and very inexpensive direct green (particularly) lasers (those cheap green lasers you find in laser pointers won’t work because they switch WAY to slow and are very unstable). Then after they spend all that R&D money they have to then sell them to you very cheap.
5) Laser Combining into each fiber. You then have the other nasty problem of getting the light from 3 lasers into a single fiber; it can be done with dichroic mirrors and the like but it has to be VERY precise or you miss the fiber. To give you some idea of the “combining” process you might want to look at my article on how Sony combined 5 lasers (2 Red, 2 Green, and 1 Blue for brightness) for a laser mirror scanning projector https://www.kguttag.com/2015/07/13/celluonsonymicrovision-optical-path/. Only now you don’t do this just once but 28 times. This problem is not impossible but requires precision and precision cost money. Maybe if you put enough R&D money into it you can make it on a single substrate. BTW, It looks to me that in the photo you see of Magic Leap prototype (https://www.wired.com/wp-content/uploads/2016/04/ff_magic_leap-eric_browy-929×697.jpg) it looks like they didn’t bother combining the lasers into single fibers.
6) Next to get the light injected into a waveguide you need to collimate the arrays of cone shaped light rays. I don’t know of any way, even with holographic optics that you can Collimate this light because you have overlapping rays of light going in different directions. You can’t collimate the individual cones of light rays or there is not way to get them to overlap to make a single image without gaps in it. I have been looking through the ML patent applications an they never seem to say how they will get this array of FSDs injected into a waveguide. You might be able to build this in a lab for one that is horribly inefficient by diffusing the light first but it would be horribly inefficient.
7) Now you have the issue of how are you going to support multiple focus planes. 72Hz is not fast enough to do it Field Sequentially so you have to put in either parallel ones so multiply by the number of focus planes. The question at this point is how much more than a Tesla Model S (starting at $66K) will it cost in production.
I think this is a big ask when you can buy an LCOS engine at 720p (and probably soon 1080p) for at about $35 per eye. The theoretical FSD advantage is that it might be able to be scaled it up to higher resolutions but you are several miracles away from that today.
There is no way to support any decent resolution with Light Fields that is going to fit on anyone’s head. It takes about 50 to 100 times the simultaneous image information to support the same resolution with a light field. Not only can’t you afford to display all the information to support good resolution, it would take and insane level of computer processing. What ML is doing is a “shortcut” of multiple focus planes which is at least possible. The “light wave display” is insane-squared, it requires the array of fibers to be in perfect sync among other issues.
ML patents show passive waveguides with multiple displays (fiber scanning or conventional) driving them. It quickly becomes cost prohibitive to support multiple displays (2 to 6 as the patents show) all with the resolution required.
Several of their figures show electrically controlled variable focus elements (VFE) optics on either side of the waveguides with one set changing the focus of a frame sequential image plane compensating while a second set of VFE compensates so the “real world” view remains in focus. There is zero probability of this working without horribly distorting the real world view.
What Magic Leap Is Highly Unlikely to Produce
Active Switching Waveguides – ML patents applications show many variations they drawn attention from other articles. The complexity of making them and the resultant cost is one big issue. There would likely be serious the degradation to the view all the layers and optical structures through to the real world. Then you have the cost both in terms of displays and optics to get images routed to the various planes of the waveguide. ML’s patent applications don’t really say how the switching would work other than saying they might use liquid crystal or lithium niobate but nothing so show they have really thought it through. I put this in the “unlikely” category because companies such as DigiLens have built switchable Bragg Gratings.