XR Fragments

Introduction How can we add more control to existing text and 3D scenes, without introducing new dataformats?
Historically, there's many attempts to create the ultimate 3D fileformat.
The lowest common denominator is: designers describing/tagging/naming things using plain text.
XR Fragments exploits the fact that all 3D models already contain such metadata: XR Fragments allows controlling of metadata in 3D scene(files) using URI's It solves:

addressibility and hypermediatic navigation of 3D scenes/objects: URI Fragments using src/href spatial metadata
Interlinking text & spatial objects by collapsing space into a Word Graph (XRWG) to show visible links
unlocking spatial potential of the (originally 2D) hashtag (which jumps to a chapter) for navigating XR documents
refraining from introducing scripting-engines for mundane tasks (and preventing its inevitable security-headaches)
the gap between text an 3d objects: object-names directly map to hashtags (=fragments), which allows 3D to text transcription.

NOTE: The chapters in this document are ordered from highlevel to lowlevel (technical) as much as possible

Core principle XR Fragments allows controlling 3D models using URLs, based on (non)existing metadata via URI's XR Fragments tries to seek to connect the world of text (semantical web / RDF), and the world of pixels.
Instead of forcing authors to combine 3D/2D objects programmatically (publishing thru a game-editor e.g.), XR Fragments integrates all which allows a universal viewing experience.
Fact: our typical browser URL's are just a possible implementation of URI's (for untapped humancentric potential of URI's see interpeer.io)

XR Fragments does not look at XR (or the web) thru the lens of HTML or URLs.
But approaches things from a higherlevel feedbackloop/hypermedia browser-perspective.

Below you can see how this translates back into good-old URLs: ?other.glb ─> #view ───> hashbus │ │ │ #filter │ │ │ │ #tag │ │ │ │ (hypermediatic) #material │ │ │ │ ( feedback ) #animation │ │ │ │ ( loop ) #texture │ │ │ │ #variable │ │ │ │ │ │ │ XRWG <─────────────────────<─────────────+ │ │ │ │ │ │ └─ objects ──────────────>─────────────+ │ │ │ │ │ +───────────────────────────────────────────────────────────────────────────────────────────────+ ]]>

?-linked and #-linked navigation are JUST one possible way to implement XR Fragments: the essential goal is to allow a Hypermediatic FeedbackLoop (HFL) between external and internal 4D navigation.

Traditional webbrowsers can become 4D document-ready by:

The XR Fragments Trinity XR Fragments utilizes URLs:

for 3D viewers/browser to manipulate the camera or objects (via URLbar)
as implicit metadata to reference (nested) objects inside 3D scene-file (local and remote)
via explicit metadata ('extras') inside 3D scene-files (interaction e.g.) or
[optionally for developers] via explicit metadata outside 3D scene-files (via sidecarfile)

List of URI Fragments

fragment	type	example	info
`#pos`	vector3	`#pos=0.5,0,0` `#pos=room` `#pos=cam2`	positions/parents camera(rig) (or XR floor) to xyz-coord/object/camera
`#rot`	vector3	`#rot=0,90,0`	rotates camera to xyz-coord 0.5,0,0
Media Fragments	media fragment	`#t=0,2&loop`	play (and loop) 3D animation from 0 seconds till 2 seconds

List of *explicit metadata These are the possible 'extras' for 3D nodes and sidecar-files

key	type	example (JSON)	function	existing compatibility
`href`	string	`"href": "b.gltf"`	XR teleport	custom property in 3D fileformats
`src`	string	`"src": "#cube"`	XR embed / teleport	custom property in 3D fileformats
`tag`	string	`"tag": "cubes geo"`	tag object (for filter-use / XRWG highlighting)	custom property in 3D fileformats
`#`	string	`"#": "#mypreset`	trigger default fragment on load	custom property in 3D fileformats

Supported popular compatible 3D fileformats: .gltf, .obj, .fbx, .usdz, .json (THREE.js), .dae and so on.

Sidecar-file

NOTE: sidecar-files break the portability of XR (Fragments) experiences, therefore side-car files are discouraged for consumer usage/sharing. However, they can accomodate developers or applications who (for whatever reason) must not modify the 3D scene-file (a .glb e.g.).

For developers, sidecar-file can allow for defining explicit XR Fragments metadata, outside of the 3D file.
This can be done via a JSON-pointers RFC6901 in a JSON sidecar-file:

experience.glb
experience.json

This would mean: hide object(s) with name or tag-value 'penguin' upon scene-load, and show it when the user clicks the chair

So after loading experience.glb the existence of experience.json is detected, to apply the explicit metadata.
The sidecar will define (or override already existing) extras, which can be handy for multi-user platforms (offer 3D scene customization/personalization to users).

In THREE.js-code this would boil down to:

Hypermediatic FeedbackLoop for XR browsers href metadata traditionally implies click AND navigate, however XR Fragments adds stateless click (xrf://#....) or navigate (xrf://#pos=...) as well (which allows many extra interactions which otherwise need a scripting language). This is known as hashbus-only events (see image above).

Being able to use the same URI Fragment DSL for navigation (href: #foo) as well as interactions (href: xrf://#bar) greatly simplifies implementation, increases HFL, and reduces need for scripting languages.

This opens up the following benefits for traditional & future webbrowsers:

hypermediatic loading/clicking 3D assets (gltf/fbx e.g.) natively (with or without using HTML).
allowing 3D assets/nodes to publish XR Fragments to themselves/eachother using the xrf:// hashbus
collapsing the 3D scene to an wordgraph (for essential navigation purposes) controllable thru a hash(tag)bus
completely bypassing the security-trap of loading external scripts (by loading 3D model-files, not HTML-javascriptable resources)

XR Fragments itself are hypermediatic and HTML-agnostic, though pseudo-XR Fragment browsers can be implemented on top of HTML/Javascript.

principle	XR 4D URL	HTML 2D URL
the XRWG	wordgraph (collapses 3D scene to tags)	Ctrl-F (find)
the hashbus	hashtags alter camera/scene/object-projections	hashtags alter document positions
src metadata	renders content and offers sourceportation	renders content
href metadata	teleports to other XR document	jumps to other HTML document
href metadata	triggers predefined view	Media fragments
href metadata	triggers camera/scene/object/projections	n/a
href metadata	draws visible connection(s) for XRWG 'tag'	n/a
href metadata	filters certain (in)visible objects	n/a
href metadata	href="xrf://#-foo&bar"	href="javascript:hideFooAndShowBar()`
	(this does not update topLevel URI)	(this is non-standard, non-hypermediatic)

An important aspect of HFL is that URI Fragments can be triggered without updating the top-level URI (default href-behaviour) thru their own 'bus' (xrf://#.....). This decoupling between navigation and interaction prevents non-standard things like (href:javascript:dosomething()).

Conventions and Definitions See appendix below in case certain terms are not clear.

XR Fragment URL Grammar For typical HTTP-like browsers/applications:

Example: ://foo.com/my3d.gltf#pos=1,0,0&prio=-5&t=0,100

Demo	Explanation
`pos=1,2,3`	vector/coordinate argument e.g.
`pos=1,2,3&rot=0,90,0&foo`	combinators

this is already implemented in all browsers

Pseudo (non-native) browser-implementations (supporting XR Fragments using HTML+JS e.g.) can use the ? search-operator to address outbound content.
In other words, the URL updates to: https://me.com?https://me.com/other.glb when navigating to https://me.com/other.glb from inside a https://me.com WebXR experience e.g.
That way, if the link gets shared, the XR Fragments implementation at https://me.com can load the latter (and still indicates which XR Fragments entrypoint-experience/client was used).

Spatial Referencing 3D XR Fragments assume the following objectname-to-URIFragment mapping:

Every 3D fileformat supports named 3D object, and this name allows URLs (fragments) to reference them (and their children objects).

Clever nested design of 3D scenes allow great ways for re-using content, and/or previewing scenes.
For example, to render a portal with a preview-version of the scene, create an 3D object with:

href: https://scene.fbx
src: https://otherworld.gltf#mainobject

It also allows sourceportation, which basically means the enduser can teleport to the original XR Document of an src embedded object, and see a visible connection to the particular embedded object. Basically an embedded link becoming an outbound link by activating it.

Level2: Implicit URI Fragments These fragments are derived from objectnames (or their extras) within a 3D scene, and trigger certain actions when evaluated by the browser:

	fragment	type	example	info
PRESET	`#<preset>`	string	`#cubes`	evaluates preset (`#foo&bar`) when a scene contains extra (`#cubes: #foo&bar` e.g.) while URL-browserbar reflects `#cubes`. Only works when metadata-key starts with `#`
FOCUS	`#<tag_or_objectname>`	string	`#person`	(and show) object(s) with `tag: person` or name `person` (XRWG lookup)
FILTERS	`#[!][-]<tag_or_objectname>[*]`	string	`#person` (`#-person`)	will reset (`!`), show/focus or hide (`-`) focus object(s) with `tag: person` or name `person` by looking up XRWG (`*`=including children)
MATERIALUPDATE	`#<tag_or_objectname>[*]=<materialname>`	string=string	`#car=metallic`	sets material of car to material with name `metallic` (`*`=including children)
			`#soldout*=halfopacity`	set material of objects tagged with `product` to material with name `metallic`
VARIABLE UPDATE	`#<variable>=<metadata-key>`	string=string	`#foo=bar`	sets URI Template variable `foo` to the value `#t=0` from existing object metadata (`bar`:`#t=0` e.g.), This allows for reactive URI Template defined in object metadata elsewhere (`src`:`://m.com/cat.mp4#{foo}` e.g., to play media using media fragment URI). NOTE: metadata-key should not start with `#`
ANIMATION	`#<tag_or_objectname>=<animationname>`	string=string	`#people=walk` `#people=noanim`	assign a different animation to object(s)

media fragments and datatypes

NOTE: below the word 'play' applies to 3D animations embedded in the 3D scene(file) but also media defined in src-metadata like audio/video-files (mp3/mp4 e.g.)

type	syntax	example	info
vector2	x,y	2,3.0	2-dimensional vector
vector3	x,y,z	2,3.0,4	3-dimensional vector
temporal W3C media fragment	t=x	0	play from 0 seconds to end (and stop)
temporal W3C media fragment	t=x,y	0,2	play from 0 seconds till 2 seconds (and stop)
temporal W3C media fragment *	s=x	1	set playback speed of audio/video/3D anim
temporal W3C media fragment *	[-]loop	loop	enable looped playback of audio/video/3D anim
		-loop	disable looped playback (does not affect playbackstate of media)
vector2	uv=u,v,uspeed,vspeed	0,0	set uv offset instantly (default speed = `1,1`)
		+0.5,+0.5	scroll instantly by adding 0.5 to the current uv coordinates
		0.2,1,0.1,0.1	scroll (lerp) to uv coordinate `0,2,1` with `0.1` units per second
		0,0,0,+0.1	scroll v coordinates with `0.1` units per second (infinitely)
		+0.5,+0.5	scroll instantly by adding 0.5 to the current uv coordinates
media parameter (shader uniform)	u:<uniform>=<string	float	vec2

* = this is extending the W3C media fragments with (missing) playback/viewport-control. Normally #t=0,2 implies setting start/stop-values AND starting playback, whereas #s=0&loop allows pausing a video, speeding up/slowing down media, as well as enabling/disabling looping. The rationale for uv is that the xywh Media Fragment deals with rectangular media, which does not translate well to 3D models (which use triangular polygons, not rectangular) positioned by uv-coordinates. This also explains the absense of a scale or rotate primitive, which is challenged by this, as well as multiple origins (mesh- or texture).

Example URI's:

https://images.org/credits.jpg#uv=0,0,0,+0.1 (infinite vertical texturescrolling)
https://video.org/organogram.mp4#t=0&loop&uv=0.1,0.1,0.3,0.3 (animated tween towards region in looped video)
https://shaders.org/plasma.glsl#t=0&u:col2=0,1,0 (red-green shader plasma starts playing from time-offset 0)

NOTE: URI Template variables are immutable and respect scope: in other words, the end-user cannot modify `blue` by entering an URL like `#blue=.....` in the browser URL, and `blue` is not accessible by the plane/media-object (however `{play}` would work). ]]>

Navigating 3D

fragment	type	functionality
<b>#pos</b>=0,0,0	vector3	position camera to 0,0,0 (+userheight in VR)
<b>#pos</b>=room	string	position camera to position of objectname `room` (+userheight in VR)
<b>#rot</b>=0,90,0	vector3	rotate camera

» example implementation
» discussion
Here's the basic level1 flow (with optional level2 features):

the Y-coordinate of pos identifies the floorposition. This means that desktop-projections usually need to add 1.5m (average person height) on top (which is done automatically by VR/AR headsets), except in case of camera-switching.
set the position of the camera accordingly to the vector3 values of #pos
if the referenced #pos object is animated, parent the current camera to that object (so it animates too)
rot sets the rotation of the camera (only for non-VR/AR headsets, however a camera-value overrules this)
level2: mediafragment t in the top-URL sets the playbackspeed and animation-range of the global scene animation
before scene load: the scene is cleared
level2: after scene load: in case the scene (rootnode) contains an # default view with a fragment value: execute non-positional fragments via the hashbus (no top-level URL change)
level2: after scene load: in case the scene (rootnode) contains an # default view with a fragment value: execute positional fragment via the hashbus + update top-level URL
level2: in case of no default # view on the scene (rootnode), default player(rig) position 0,0,0 is assumed.
in case a href does not mention any pos-coordinate, the current position will be assumed

Here's an ascii representation of a 3D scene-graph which contains 3D objects ◻ and their metadata: An XR Fragment-compatible browser viewing this scene, allows the end-user to interact with the buttonA and buttonB.
In case of buttonA the end-user will be teleported to another location and time in the current loaded scene, but buttonB will replace the current scene with a new one, like other.fbx, and assume pos=0,0,0.

Top-level URL processing

Example URL: ://foo/world.gltf#cube&pos=0,0,0

The URL-processing-flow for hypermedia browsers goes like this:

IF a #cube matches a custom property-key (of an object) in the 3D file/scene (#cube: #......) THEN execute that predefined_view.
IF scene operators (pos) and/or animation operator (t) are present in the URL then (re)position the camera and/or animation-range accordingly.
IF no camera-position has been set in step 1 or 2 update the top-level URL with #pos=0,0,0 (example)
IF a #cube matches the name (of an object) in the 3D file/scene then draw a line from the enduser('s heart) to that object (to highlight it).
IF a #cube matches anything else in the XR Word Graph (XRWG) draw wires to them (text or related objects).

Embedding XR content using src src is the 3D version of the <a target="_blank" href="https://www.w3.org/html/wiki/Elements/iframe">iframe</a>.
It instances content (in objects) in the current scene/asset, and follows similar logic like the previous chapter, except that it does not modify the camera.

fragment	type	example value
`src`	string (uri, hashtag/filter)	`#cube` `#sometag` #cube&-ball_inside_cube`<br>`#-sky&-rain`<br>`#-language&english`<br>`#price=>5`<br>`https://linux.org/penguin.png` `https://linux.world/distrowatch.gltf#t=1,100` `linuxapp://conference/nixworkshop/apply.gltf#-cta&cta_apply` `androidapp://page1?tutorial#pos=0,0,1&t1,100` `foo.mp3#0,0,0`

Here's an ascii representation of a 3D scene-graph with 3D objects ◻ which embeds remote & local 3D objects ◻ with/out using filters: An XR Fragment-compatible browser viewing this scene, lazy-loads and projects painting.png onto the (plane) object called canvas (which is copy-instanced in the bed and livingroom).
Also, after lazy-loading ocean.com/aquarium.gltf, only the queried objects fishbowl (and bass and tuna) will be instanced inside aquariumcube.
Resizing will be happen accordingly to its placeholder object aquariumcube, see chapter Scaling.

Instead of cherrypicking a rootobject #fishbowl with src, additional filters can be used to include/exclude certain objects. See next chapter on filtering below.

Specification:

local/remote content is instanced by the src (filter) value (and attaches it to the placeholder mesh containing the src property)
by default all objects are loaded into the instanced src (scene) object (but not shown yet)
local src values (#... e.g.) starting with a non-negating filter (#cube e.g.) will (deep)reparent that object (with name cube) as the new root of the scene at position 0,0,0
local src values should respect (negative) filters (#-foo&price=>3)
the instanced scene (from a src value) should be scaled accordingly to its placeholder object or scaled relatively based on the scale-property (of a geometry-less placeholder, an 'empty'-object in blender e.g.). For more info see Chapter Scaling.
external src values should be served with appropriate mimetype (so the XR Fragment-compatible browser will now how to render it). The bare minimum supported mimetypes are:
src values should make its placeholder object invisible, and only flush its children when the resolved content can succesfully be retrieved (see broken links)
external src values should respect the fallback link mechanism (see broken links
when the placeholder object is a 2D plane, but the mimetype is 3D, then render the spatial content on that plane via a stencil buffer.
src-values are non-recursive: when linking to an external object (src: foo.fbx#bar), then src-metadata on object bar should be ignored.
an external src-value should always allow a sourceportation icon within 3 meter: teleporting to the origin URI to which the object belongs.
when only one object was cherrypicked (#cube e.g.), set its position to 0,0,0
when the enduser clicks an href with #t=1,0,0 (play) will be applied to all src mediacontent with a timeline (mp4/mp3 e.g.)
a non-euclidian portal can be rendered for flat 3D objects (using stencil buffer e.g.) in case ofspatial src-values (an object #world3 or URL world3.fbx e.g.).

model/gltf-binary
model/gltf+json
image/png
image/jpg
text/plain;charset=utf-8

» example implementation
» example 3D asset
» discussion

Navigating content href portals navigation, portals & mutations

fragment	type	example value
`href`	string (uri or predefined view)	`#pos=1,1,0` `#pos=1,1,0&rot=90,0,0` `://somefile.gltf#pos=1,1,0`

clicking an outbound ''external''- or ''file URI'' fully replaces the current scene and assumes pos=0,0,0&rot=0,0,0 by default (unless specified)
relocation/reorientation should happen locally for local URI's (#pos=....)
navigation should not happen ''immediately'' when user is more than 5 meter away from the portal/object containing the href (to prevent accidental navigation e.g.)
URL navigation should always be reflected in the client URL-bar (in case of javascript: see [here for an example navigator), and only update the URL-bar after the scene (default fragment #) has been loaded.
In immersive XR mode, the navigator back/forward-buttons should be always visible (using a wearable e.g., see [here for an example wearable)
make sure that the ''back-button'' of the ''browser-history'' always refers to the previous position (see [here)
ignore previous rule in special cases, like clicking an href using camera-portal collision (the back-button could cause a teleport-loop if the previous position is too close)
href-events should bubble upward the node-tree (from children to ancestors, so that ancestors can also contain an href), however only 1 href can be executed at the same time.
the end-user navigator back/forward buttons should repeat a back/forward action until a pos=... primitive is found (the stateless xrf:// href-values should not be pushed to the url-history)

» example implementation
» example 3D asset
» discussion

Walking surfaces XR Fragment-compatible viewers can infer this data based scanning the scene for:

materialless (nameless & textureless) mesh-objects (without src and href)

optionally the viewer can offer thumbstick, mouse or joystick teleport-tools for non-roomscale VR/AR setups.

UX spec End-users should always have read/write access to:

the current (toplevel) URL (an URLbar etc)
URL-history (a back/forward button e.g.)
Clicking/Touching an href navigates (and updates the URL) to another scene/file (and coordinate e.g. in case the URL contains XR Fragments).

Scaling instanced content Sometimes embedded properties (like src) instance new objects.
But what about their scale?
How does the scale of the object (with the embedded properties) impact the scale of the referenced content?

Rule of thumb: visible placeholder objects act as a '3D canvas' for the referenced scene (a plane acts like a 2D canvas for images e, a cube as a 3D canvas e.g.).

IF an embedded property (src e.g.) is set on an non-empty placeholder object (geometry of >2 vertices):

calculate the bounding box of the ''placeholder'' object (maxsize=1.4 e.g.)
hide the ''placeholder'' object (material e.g.)
instance the src scene as a child of the existing object
calculate the bounding box of the instanced scene, and scale it accordingly (to 1.4 e.g.)

REASON: non-empty placeholder object can act as a protective bounding-box (for remote content of which might grow over time e.g.)

ELSE multiply the scale-vector of the instanced scene with the scale-vector (a common property of a 3D node) of the placeholder object.

TODO: needs intermediate visuals to make things more obvious

XR Fragment: pos [[» example implementation|https://github.com/coderofsalvation/xrfragment/blob/main/src/3rd/js/three/xrf/pos.js]]

XR Fragment: rot [[» example implementation|https://github.com/coderofsalvation/xrfragment/blob/main/src/3rd/js/three/xrf/pos.js]]

XR Fragment: t [[» example implementation|https://github.com/coderofsalvation/xrfragment/blob/main/src/3rd/js/three/xrf/t.js]]

XR audio/video integration To play global audio/video items:

add a src: foo.mp3 or src: bar.mp4 metadata to a 3D object (cube e.g.)
to enable auto-play and global timeline ([[#t=|t]]) control: hardcode a [[#t=|t]] XR Fragment: (src: bar.mp3#t=0&loop e.g.)
to play it, add href: #cube somewhere else
to enable enduser-triggered play, use a [[URI Template]] XR Fragment: (src: bar.mp3#{player} and play: t=0&loop and href: xrf://#player=play e.g.)
when the enduser clicks the href, #t=0&loop (play) will be applied to the src value

NOTE: hardcoded framestart/framestop uses sampleRate/fps of embedded audio/video, otherwise the global fps applies. For more info see [[#t|t]].

XR Fragment filters Include, exclude, hide/shows objects using space-separated strings:

example	outcome
`#-sky`	show everything except object named `sky`
`#-language&english`	hide everything with tag `language`, but show all tag `english` objects
`#-price&price=>10`	hide all objects with property `price`, then only show object with price above 10
`#-house*`	hide `house` object and everything inside (=`*`)

It's simple but powerful syntax which allows filtering the scene using searchengine prompt-style feeling:

filters are a way to traverse a scene, and filter objects based on their name, tag- or property-values.

see an (outdated) example video here which used a dedicated q= variable (now deprecated and usable directly)

including/excluding By default, selectors work like photoshop-layers: they scan for matching layer(name/properties) within the scene-graph. Each matched object (not their children) will be toggled (in)visible when selecting.

operator	info
`-`	hides object(s) (`#-myobject&-objects` e.g.
`=`	indicates an object-embedded custom property key/value (`#price=4&category=foo` e.g.)
`=>` `=<`	compare float or int number (`#price=>4` e.g.)
`*`	deepselect: automatically select children of selected object, including local (nonremote) embedded objects (starting with `#`)

NOTE 1: after an external embedded object has been instanced (src: https://y.com/bar.fbx#room e.g.), filters do not affect them anymore (reason: local tag/name collisions can be mitigated easily, but not in case of remote content). NOTE 2: depending on the used 3D framework, toggling objects (in)visible should happen by enabling/disableing writing to the colorbuffer (to allow children being still visible while their parents are invisible).

» example implementation » example 3D asset » discussion

Filter Parser Here's how to write a filter parser:

create an associative array/object to store filter-arguments as objects
detect object id's & properties foo=1 and foo (reference regex= ~/^.*=[><=]?/ )
detect excluders like -foo,-foo=1,-.foo,-/foo (reference regex= /^-/ )
detect root selectors like /foo (reference regex= /^[-]?\// )
detect number values like foo=1 (reference regex= /^[0-9\.]+$/ )
detect operators so you can easily strip keys (reference regex= /(^-|\*$)/ )
detect exclude keys like -foo (reference regex= /^-/ )
for every filter token split string on =
and we set root to true or false (true=/ root selector is present)
therefore we we set show to true or false (false=excluder -)

An example filter-parser (which compiles to many languages) can be found here

Visible links When predefined views, XRWG fragments and ID fragments (#cube or #mytag e.g.) are triggered by the enduser (via toplevel URL or clicking href):

draw a wire from the enduser (preferabbly a bit below the camera, heartposition) to object(s) matching that ID (objectname)
draw a wire from the enduser (preferabbly a bit below the camera, heartposition) to object(s) matching that tag value
draw a wire from the enduser (preferabbly a bit below the camera, heartposition) to object(s) containing that in their src or href value

The obvious approach for this, is to consult the XRWG (example), which basically has all these things already collected/organized for you during scene-load. UX

do not update the wires when the enduser moves, leave them as is
offer a control near the back/forward button which allows the user to (turn off) control the correlation-intensity of the XRWG

Text in XR (tagging,linking to spatial objects) How does XR Fragments interlink text with objects?

The XR Fragments does this by collapsing space into a Word Graph (the XRWG example), augmented by Bib(s)Tex.

Instead of just throwing together all kinds media types into one experience (games), what about their tagged/semantical relationships?
Perhaps the following question is related: why is HTML adopted less in games outside the browser? Hence:

XR Fragments promotes (de)serializing a scene to a (lowercase) XRWG (example)
XR Fragments primes the XRWG, by collecting words from the tag and name-property of 3D objects.
XR Fragments primes the XRWG, by collecting words from optional metadata at the end of content of text (see default mimetype & Data URI)
XR Fragments primes the XRWG, by collecting tags/id's from linked hypermedia (URI fragments for HTML e.g.)
The XRWG should be recalculated when textvalues (in src) change
HTML/RDF/JSON is still great, but is beyond the XRWG-scope (they fit better in the application-layer, or as embedded src content)
Applications don't have to be able to access the XRWG programmatically, as they can easily generate one themselves by traversing the scene-nodes.
The XR Fragment focuses on fast and easy-to-generate end-user controllable word graphs (instead of complex implementations that try to defeat word ambiguity)
Instead of exact lowercase word-matching, levensteihn-distance-based matching is preferred

Example of generating XRWG out of the XRWG and textdata with hashtags: This allows hasslefree authoring and copy-paste of associations for and by humans, but also makes these URLs possible:

URL example	Result
`https://my.com/foo.gltf#baroque`	draws lines between 3D mesh `castle`, and `mydoc`'s text `baroque`
`https://my.com/foo.gltf#john`	draws lines between mesh `john`, and the text `John` of `mydoc`
`https://my.com/foo.gltf#house`	draws lines between mesh `castle`, and other objects with tag `house` or `todo`

the URI fragment #john&mydoc&house would draw a connection between these 3 meshes.

The XRWG allows endusers to show/hide relationships in realtime in XR Browsers at various levels:

wordmatch inside src text
wordmatch inside href text
wordmatch object-names
wordmatch object-tagnames

Spatial wires can be rendered between words/objects etc.
Some pointers for good UX (but not necessary to be XR Fragment compatible):

The XR Browser needs to adjust tag-scope based on the endusers needs/focus (infinite tagging only makes sense when environment is scaled down significantly)
The XR Browser should always allow the human to view/edit the metadata, by clicking 'toggle metadata' on the 'back' (contextmenu e.g.) of any XR text, anywhere anytime.
respect multi-line BiBTeX metadata in text because of the core principle
Default font (unless specified otherwise) is a modern monospace font, for maximized tabular expressiveness (see the core principle).
anti-pattern: hardcoupling an XR Browser with a mandatory markup/scripting-language which departs from onubtrusive plain text (HTML/VRML/Javascript) (see the core principle)
anti-pattern: limiting human introspection, by abandoning plain text as first tag citizen.

Default Data URI mimetype The src-values work as expected (respecting mime-types), however: The XR Fragment specification advices to bump the traditional default browser-mimetype text/plain;charset=US-ASCII to a hashtag-friendly one: text/plain;charset=utf-8;hashtag This indicates that:

utf-8 is supported by default
words beginning with # (hashtags) will prime the XRWG by adding the hashtag to the XRWG, linking to the current sentence/paragraph/alltext (depending on '.') to the XRWG

Advantages:

out-of-the-box (de)multiplex human text and metadata in one go (see the core principle)
no network-overhead for metadata (see the core principle)
ensuring high FPS: realtime HTML/RDF historically is too 'requesty'/'parsy' for game studios
rich send/receive/copy-paste everywhere by default, metadata being retained (see the core principle)
netto result: less webservices, therefore less servers, and overall better FPS in XR

This significantly expands expressiveness and portability of human tagged text, by postponing machine-concerns to the end of the human text in contrast to literal interweaving of content and markupsymbols (or extra network requests, webservices e.g.).

For all other purposes, regular mimetypes can be used (but are not required by the spec).

URL and Data URI The enduser will only see welcome human and Hello friends rendered verbatim (see mimetype). The beauty is that text in Data URI automatically promotes rich copy-paste (retaining metadata). In both cases, the text gets rendered immediately (onto a plane geometry, hence the name '_canvas'). The XR Fragment-compatible browser can let the enduser access visual-meta(data)-fields after interacting with the object (contextmenu e.g.).

additional tagging using bibs: to tag spatial object note_canvas with 'todo', the enduser can type or speak #note_canvas@todo

Importing/exporting For usecases like importing/exporting/p2p casting a scene, the issue of external files comes into play.

export: if the 3D scene contains relative src/href values, rewrite them into absolute URL values.

Reflection Mapping Environment mapping is crucial for creating realistic reflections and lighting effects on 3D objects. To apply environment mapping efficiently in a 3D scene, traverse the scene graph and assign each object's environment map based on the nearest ancestor's texture map. This ensures that objects inherit the correct environment mapping from their closest parent with a texture, enhancing the visual consistency and realism. Most 3D viewers apply one and the same environment map for various models, however this logic allows a more natural & automatic strategy for reflection mapping:

traverse the scene graph depth-first
remember the most recent parentnode (P) with a texture material
for every non-root node with a texture material 3.1 clone that material (as materials might be shared across objects) 3.2 set the environmentmap to the last known parent texture (P)

Transclusion (broken link) resolution In spirit of Ted Nelson's 'transclusion resolution', there's a soft-mechanism to harden links & minimize broken links in various ways:

defining a different transport protocol (https vs ipfs or DAT) in src or href values can make a difference
mirroring files on another protocol using (HTTP) errorcode tags in src or href properties
in case of src: nesting a copy of the embedded object in the placeholder object (embeddedObject) will not be replaced when the request fails

due to the popularity, maturity and extensiveness of HTTP codes for client/server communication, non-HTTP protocols easily map to HTTP codes (ipfs ERR_NOT_FOUND maps to 404 e.g.)

For example:

Topic-based index-less Webrings As hashtags in URLs map to the XWRG, href-values can be used to promote topic-based index-less webrings.
Consider 3D scenes linking to eachother using these href values:

href: schoolA.edu/projects.gltf#math
href: schoolB.edu/projects.gltf#math
href: university.edu/projects.gltf#math

These links would all show visible links to math-tagged objects in the scene.
To filter out non-related objects one could take it a step further using filters:

href: schoolA.edu/projects.gltf#math&-topics math
href: schoolB.edu/projects.gltf#math&-courses math
href: university.edu/projects.gltf#math&-theme math

This would hide all object tagged with topic, courses or theme (including math) so that later only objects tagged with math will be visible

This makes spatial content multi-purpose, without the need to separate content into separate files, or show/hide things using a complex logiclayer like javascript.

URI Templates (RFC6570) XR Fragments adopts Level1 URI Fragment expansion to provide safe interactivity.
The following demonstrates a simple video player:

Additional scene metadata XR Fragments does not aim to redefine the metadata-space or accessibility-space by introducing its own cataloging-metadata fields. Instead, it encourages browsers to scan nodes for the following custom properties:

SPDX license information
ARIA attributes (aria-*: .....)
Open Graph attributes (og:*: .....)
Dublin-Core attributes(dc:*: .....)
BibTex when known bibtex-keys exist with values enclosed in { and },

ARIA (aria-description) is the most important to support, as it promotes accessibility and allows scene transcripts. Please start aria-description with a verb to aid transcripts.

Example: object 'tryceratops' with aria-description: is a huge dinosaurus standing on a #mountain generates transcript #tryceratops is a huge dinosaurus standing on a #mountain, where the hashtags are clickable XR Fragments (activating the visible-links in the XR browser).

Individual nodes can be enriched with such metadata, but most importantly the scene node:

metadata key	example value
`aria-description`, `og:description`, `dc:description`	`An immersive experience about Triceratops` (*)
`SPDX`	`CC0-1.0`
`dc:creator`	`John Doe`
`dc:title`, `og:title`	'Triceratops` (*)
`og:site_name`	`https://xrfragment.org`
`dc.publisher`	`NLNET`
`dc.date`	`2024-01-01`
`dc.identifier`	`XRFRAGMENT-001`
`journal` (bibTeX)	`{Future Of Text Vol 3},`

* = these are interchangable (only one needs to be defined)

There's no silver bullet when it comes to metadata, so one should support where the metadata is/goes.

These attributes can be scanned and presented during an href or src eye/mouse-over.

Accessibility interface The addressibility of XR Fragments allows for unique 3D-to-text transcripts, as well as an textual interface to navigate 3D content.
Spec:

The enduser must be able to enable an accessibility-mode (which persists across application/webpage restarts)
Accessibility-mode must contain a text-input for the user to enter text
Accessibility-mode must contain a flexible textlog for the user to read (via screenreader, screen, or TTS e.g.)
the textlog contains aria-descriptions, and its narration (Screenreader e.g.) can be skipped (via 2-button navigation)
The back command should navigate back to the previous URL (alias for browser-backbutton)
The forward command should navigate back to the next URL (alias for browser-nextbutton)
A destination is a 3D node containing an href with a pos= XR fragment
The go command should list all possible destinations
The go left command should move the camera around 0.3 meters to the left
The go right command should move the camera around 0.3 meters to the right
The go forward command should move the camera 0.3 meters forward (direction of current rotation).
The rotate left command should rotate the camera 0.3 to the left
The rotate left command should rotate the camera 0.3 to the right
The (dynamic) go abc command should navigate to #pos=scene2 in case there's a 3D node with name abc and href value #pos=scene2
The look command should give an (contextual) 3D-to-text transcript, by scanning the aria-description values of the current pos= value (including its children)
The do command should list all possible href values which don't contain an pos= XR Fragment
The (dynamic) do abc command should navigate/execute https://.../... in case a 3D node exist with name abc and href value https://.../...

Two-button navigation For specific user-profiles, gyroscope/mouse/keyboard/audio/visuals will not be available.
Therefore a 2-button navigation-interface is the bare minimum interface:

objects with href metadata can be cycled via a key (tab on a keyboard)
objects with href metadata can be activated via a key (enter on a keyboard)
the TTS reads the href-value (and/or aria-description if available)

Overlap with fileformat-specific extensions Some 3D scene-fileformats have support for extensions. What if the functionality of those overlap? For example, GLTF has the OMI_LINK extension which might overlap with XR Fragment's href:

Priority Order and Precedence, otherwise fallback applies

1.Extensions Take Precedence: Since glTF-specific extensions are designed with the format’s specific needs and optimizations in mind, they should take precedence over extras metadata in cases where both contain overlapping functionality. This approach aligns with the idea that extensions are more likely to be interpreted uniformly by glTF-compatible software.

Fallback Fall-through Mechanism: If a glTF implementation does not support a particular extension, the (XRF) extras field can serve as a fallback. This way, metadata provided in extras can still be useful for applications that don't handle certain extensions.

Example 1 In case of the OMI_LINK glTF extension (href: https://nlnet.nl) and an XR Fragment (href: #pos=otherroom or href: otherplanet.glb), it is clear that https://nlnet.nl should open in a browsertab, whereas the XR Fragment links should teleport the user. If the OMI_LINK contains an XR Fragment (#pos= e.g.) a teleport should be performed only (and other [overlapping] metadata should be ignored). Example 2 If an Extensions uses XR Fragments in URI's (href: #pos=otherroom or href: xrf://-walls in OMI_LINK e.g.), then perform them according to XR Fragment spec (teleport user). But only once: ignore further overlapping metadata for that usecase.

Vendor Prefixes Vendor-specific metadata in a 3D scenefiles, are similar to vendor-specific CSS-prefixes (-moz-opacity: 0.2 e.g.). This allows popular 3D engines/frameworks, to initialize specific features when loading a scene/object, in a progressive enhanced way. Vendor Prefixes allows embedding 3D engines/framework-specific features a 3D file via metadata:

what	XR metadata	Lowest common denominator
CSS	vendor-agnostic	2D canvas + object referencing/styling
XR Fragments	vendor-agnostic	3D camera + object(file) load/embed/click/referencing
Vendor prefixs	vendor-specific	Specialized Entity-Component implementation

Why? Because not all XR interactions can/should be solved/standardized by embedding XR Fragments into any 3D file. The lowest common denominator between 3D engines is the 'entity'-part of their entity-component-system (ECS). The 'component'-part can be progressively enhanced via vendor prefixes.

For example, the following metadata can be added to a .glb file, to make an object grabbable in AFRAME: inits 'grabbable' component on object john │ │ | / \ ├─ -aframe-material.color: '#F0A' | > inits 'material' component on object john │ │ | ├─ -aframe-text.value: '{name}{age}'| > inits 'text' component (*) with value 'john' │ │ | ├─ -three-material.fog: false | > changes material settings in THREE.js app │ │ | ├─ -godot-Label3D.text: '{name}{age}'| > inits 'Label3D' component (*) in Godot │ │ +--------+ | │ │ | │ ├─ -GODOT-version: '4.3' | > exporters/authors can report targeted version │ ├─ -AFRAME-version: '1.6.0' | and (optionally) hint component-repo│ ├─ -AFRAME-info: 'https://git.benetou.fr/comps' │ │ | │ +────────────────────────────────────────────────────────────────────────────────────────────────────────+ ]]>

key/value syntax: -<vendorname>-<component|version>.<key> [string/boolean/float/int]-value

String-templatevalues are evaluated as per URI Templates (RFC6570) Level 1.

This 'separating of mechanism from policy' (unix rule) does somewhat break portability of an XR experience, but still prevents (E-waste of) handcoded virtual worlds. It allows for (XR experience) metadata to survive in future 3D engines and scene-fileformats.

Security Considerations The only dynamic parts are W3C Media Fragments and URI Templates (RFC6570).
The use of URI Templates is limited to pre-defined variables and Level0 fragments-expansion only, which makes it quite safe.
n fact, it is much safer than relying on a scripting language (javascript) which can change URN too.

FAQ Q: Why is everything HTTP GET-based, what about POST/PUT/DELETE HATEOS
A: Because it's out of scope: XR Fragment specifies a read-only way to surf XR documents. These things belong in the application layer (for example, an XR Hypermedia browser can decide to support POST/PUT/DELETE requests for embedded HTML thru src values) Q: Why isn't there support for scripting, URI Template Fragments are so limited compared to WASM & javascript A: This is out of scope as it unhyperifies hypermedia, and this is up to XR hypermedia browser-extensions.
Historically scripting/Javascript seems to been able to turn webpages from hypermedia documents into its opposite (hyperscripted nonhypermedia documents).
In order to prevent this backward-movement (hypermedia tends to liberate people from finnicky scripting) XR Fragment uses W3C Media Fragments and URI Templates (RFC6570), to prevent unhyperifying itself by hardcoupling to a particular markup or scripting language.
XR Fragments supports filtering objects in a scene only, because in the history of the javascript-powered web, showing/hiding document-entities seems to be one of the most popular basic usecases.
Doing advanced scripting & networkrequests under the hood are obviously interesting endavours, but this is something which should not be hardcoupled with XR Fragments or hypermedia.
This perhaps belongs more to browser extensions.
Non-HTML Hypermedia browsers should make browser extensions the right place, to 'extend' experiences, in contrast to code/javascript inside hypermedia documents (this turned out as a hypermedia antipattern).

authors

Leon van Kammen (@lvk@mastodon.online)
Jens Finkhäuser (@jens@social.finkhaeuser.de)

IANA Considerations This document has no IANA actions.

Acknowledgments

NLNET
Future of Text
visual-meta.info
Michiel Leenaars
Gerben van der Broeke
Mauve
Jens Finkhäuser
Marc Belmont
Tim Gerritsen
Frode Hegland
Brandel Zackernuk
Mark Anderson

Appendix: Definitions

definition	explanation
human	a sentient being who thinks fuzzy, absorbs, and shares thought (by plain text, not markuplanguage)
scene	a (local/remote) 3D scene or 3D file (index.gltf e.g.)
3D object	an object inside a scene characterized by vertex-, face- and customproperty data.
URI	some resource at something somewhere via someprotocol (`http://me.com/foo.glb#foo` or `e76f8efec8efce98e6f` see interpeer.io)
URL	something somewhere via someprotocol (`http://me.com/foo.glb`)
URN	something at some domain (`me.com/foo.glb`)
metadata	custom properties of text, 3D Scene or Object(nodes), relevant to machines and a human minority (academics/developers)
XR fragment	URI Fragment with spatial hints like `#pos=0,0,0&t=1,100` e.g.
the XRWG	wordgraph (collapses 3D scene to tags)
the hashbus	hashtags map to camera/scene-projections
spacetime hashtags	positions camera, triggers scene-preset/time
teleportation	repositioning the enduser to a different position (or 3D scene/file)
sourceportation	teleporting the enduser to the original XR Document of an `src` embedded object.
placeholder object	a 3D object which with src-metadata (which will be replaced by the src-data.)
src	(HTML-piggybacked) metadata of a 3D object which instances content
href	(HTML-piggybacked) metadata of a 3D object which links to content
filter	URI Fragment(s) which show/hide object(s) in a scene based on name/tag/property (`#cube&-price=>3`)
visual-meta	visual-meta data appended to text/books/papers which is indirectly visible/editable in XR.
requestless metadata	metadata which never spawns new requests (unlike RDF/HTML, which can cause framerate-dropping, hence not used a lot in games)
FPS	frames per second in spatial experiences (games,VR,AR e.g.), should be as high as possible
introspective	inward sensemaking ("I feel this belongs to that")
extrospective	outward sensemaking ("I'm fairly sure John is a person who lives in oklahoma")
`◻`	ascii representation of an 3D object/mesh
(un)obtrusive	obtrusive: wrapping human text/thought in XML/HTML/JSON obfuscates human text into a salad of machine-symbols and words
flat 3D object	a 3D object of which all verticies share a plane
BibTeX	simple tagging/citing/referencing standard for plaintext
BibTag	a BibTeX tag
(hashtag)bibs	an easy to speak/type/scan tagging SDL (see here which expands to BibTex/JSON/XML