The moment a new model is signed off, what usually holds up progress is not R&D but "getting everyone who needs to know it to know it". Head office schedules a central training session and dealers fly in from all over; a salesperson who misses that round waits for next quarter; overseas customers, separated by time zones, get a recording. By the time everyone has finally been through it once, the machine needs a revision.

This article does not pitch "using 3D for training". It takes the problem apart: which parts of training are genuinely stuck, what the mechanism-level cost of the old approaches is, and how a 3D training space deployed on your own server, entered by opening a browser works in practice — what to prepare, and what it cannot do.

1. These kinds of training are the ones that get stuck

Not every kind of training is worth moving into 3D. The genuinely awkward ones are these:

• A new model or version ships and the channel must keep up: dealers, distributor outlets and after-sales outlets nationwide all have to learn it inside the same window; anyone a step behind may install it wrong or describe it wrong;
• Equipment operation and safety procedures for new hires: they read the manual first, nobody dares let them loose on the real machine, and mentors cannot fit everyone into their schedules;
• After-sales fault handling: the same class of fault is asked about repeatedly by different customers, and engineers explain it again and again, burning effort on repetition;
• Overseas and remote sites: time zones, language, visas, travel — remote guidance ends up as gestures over the phone;
• Assessment drills: people need to practise, but not on the real machine — especially anything live, pressurised or dangerous;
• Machines too large or too expensive: a sample unit cannot be shipped, and only a couple of training rigs can be bought, so dozens of people queue up for them.

The common thread in one sentence: the people who need to know are scattered everywhere, and there is only one machine.

2. The mechanism-level cost of the old approaches

Path A: Gather everyone at head office

It is thorough, questions get answered on the spot, and people get hands-on. The cost is scheduling and travel: only a few sessions a year, each covering a limited number of people; dealers close for several days to attend; anyone who misses it waits for the next round; and when the machine is revised, the slides and the recording have to be redone.

Path B: Manual + recorded video + phone support

Low cost, and it can be sent to everyone. But the mechanism is one-way: learners can only see the angle you chose, and cannot turn to the side where they are stuck; questions like assembly order — which part comes off first, where the clips are — are far better answered by turning the thing around than by hearing it explained ten times on video; and every revision means re-cutting the video.

Path C: Ship sample units / buy training rigs

Close to the real-machine feel. But a sample unit carries freight, depreciation and damage risk; only a couple of training rigs can be bought, so dozens of people queue; and it can only sit in one place, so dealers elsewhere still cannot reach it.

Path D: A centralised SaaS training platform

You can view 3D in a browser. But content and learner data are hosted on someone else's servers, which is hard to justify if training material is not supposed to circulate; the training flow is constrained by templates, so following your own machine's real disassembly order often cannot be done; and it is billed annually — stop paying and it goes offline, taking years of courseware and records with it.

What all four share: either they do not go deep enough, or they cost too much, or the control is not in your hands.

3. The four paths side by side

4. What putting the training space on your own server actually looks like

Genesis (the Genesis Virtual World CRM System) is not a finished set of courseware, nor a centralised platform, but a browser-side, self-hostable Three.js 3D virtual world foundation / framework. Install it on your own server and you get a world address that belongs to you.

Using it as a training space, the mechanism-level differences are:

• Open a browser and you are in — no software, no headset: dealers click the link on a desktop, phone or tablet. This is the key step for "reaching remote people";
• One machine becomes a walkable scene: put the machine's GLB model into the space, split and place parts according to the real structure, and where needed attach disassembly sequences and cautions to specific parts (the Three.js foundation supports secondary development defined around your own product, rather than filling in a template);
• Multiplayer, with instructor and learners in the same space: not watching the instructor share a screen, but a dozen avatars standing beside the same machine while the instructor walks over and points at what he is explaining;
• Private data: the system is deployed on your own server; machine structure material, learner lists and training records do not pass through a third-party platform's servers. For manufacturers with requirements around drawings and material circulating, this is often the decision-critical point;
• Assets reusable long term: when the machine is revised, update the model and the point notes rather than tearing down the whole course; and there is no going offline when payment stops — as long as your server is up, the world is there.

5. If you add an "off-shift never ends" docent (an existing capability)

In 2026 the system finished its AI access capability: an AI can walk into the 3D world as a human-shaped character — with a body, with coordinates, seen in the scene by real people, able to speak, walk and lead the way.

Applied to training, it can be used like this:

• Write an "AI description" for each training point (filled in directly in the editor backend, no more than 500 characters) telling the AI what is here and what to explain;
• When a learner comes in to practise on their own, the AI docent walks over, explains according to the point and answers questions;
• The business side connects its own knowledge base (fault trees, common Q&A, manual text) so answers come from its own material;
• The AI's identity is disclosed: when it enters, the system posts an "(AI) joined" notice and the name carries an AI prefix — real people always know the other side is an AI.

It has to be said clearly that this AI capability has hard boundaries:

• the platform does not host a knowledge base; industry material is connected locally by the business itself;
• the platform does no speech recognition or synthesis, voice relay is off by default, and "able to listen and speak" has to be carried by the business's own AI client;
• the AI cannot see the picture; what it receives is a structured "spatial radar" (who is nearby, how far, facing which way, which objects are around), not a camera feed;
• the AI does no terrain conforming; movement is a plane with boundaries;
• the AI cannot teleport and cannot touch assets, following the same rules as a human visitor.

In other words, it is a programmable character whose behaviour comes from the model and the prompts you give it, not something with a mind of its own.

6. What to prepare for deployment

Listed by usable version — no need to do it all at once:

1. A server + a domain name: deploy the system and get your own world address (running it on a local network or a single machine first is fine);
2. Machine models: export GLB. If detailed models are not available yet, run the whole flow through with simplified ones and replace them gradually;
3. A training script: which machines, how many steps per machine, what to explain at each point — this script determines the skeleton of the 3D space;
4. Point notes: if you use the AI docent, add a description of no more than 500 characters for each point;
5. Knowledge base material: fault trees, common Q&A, manual text, organised as retrievable content on the business side;
6. Someone who can modify the front end: uploading models and placing points needs no coding; deep interaction (step checking, scored assessment) means secondary development on the Three.js foundation;
7. An AI access environment: if you use the docent, you need an environment that runs Node 18+, plus an access key. When the key is generated on the platform the plain text is shown only once, so save it on the spot.

The iteration rhythm of the training content changes as well: when the machine is revised, update the model and the point notes first, and tomorrow dealers see the new version the moment they open the link — no waiting for the next central session.

7. What this approach is not suited to

There are scenarios where I will not claim it fits:

• Anyone who needs hands-on real-machine feel: the torque when tightening a bolt, judging an anomaly by sound — a 3D world cannot give those. It only replaces the layer of "understanding the structure and remembering the sequence";
• Anyone expecting the AI to read the fine detail of machine images on its own: the current version has no visual understanding capability and cannot do inspection-type work;
• Anyone wanting the AI to act in the physical world: it exists only inside this 3D world;
• Anyone wanting the AI to make high-risk decisions autonomously: we do not do it, and would not recommend it.

One more: if your company has nobody who touches code and no plan to bring someone in for the deployment, the startup cost of this will be higher than you expect — it is a foundation, not an out-of-the-box courseware product.

8. FAQ

Q: Do dealers need to install software or buy a headset?

A: No. It is browser-based — desktop, phone or tablet, open the link and you are in the 3D space. No client download, no headset.

Q: Our machine models are confidential. Will putting them in a 3D world leak them?

A: The system is deployed on your own server; models and training data do not pass through a third-party platform's servers, and there is no going offline when payment stops. How material circulates is controlled by your own deployment environment.

Q: We have no 3D models. Can we still do this?

A: Start with simplified models to get the training flow working — settle "which machines, which steps, what to say at each point" first, then replace them with detailed models gradually. The flow itself is often worth more than model precision.

Q: Can it record who has trained and who has not?

A: The system includes an admin backend and a user system, usable for managing who enters the world and their permissions, and training records can land on your own server. The exact fields depend on the version you deploy.

Q: If the instructor is away and learners come in on their own, is anyone there to answer?

A: An AI docent can be configured. It answers from the point descriptions you give it and the business-side knowledge base, and walks the scene as a human-shaped character. Note, though, that it cannot see the picture and does no speech recognition or synthesis, and cannot take on inspection-type work that requires judging images.

Q: When the machine is revised later, does the whole thing have to be redone?

A: No. Update the model and the point notes; dealers see the latest version the moment they open the link, without waiting for the next central session.

9. Source Code and Repositories

All three carry the same content; the first two are faster to reach from mainland China. The repositories contain the deployment guide and a demo entry point.

• Gitee (faster from mainland China): https://gitee.com/miduoxinxijeji/miduo.git
• GitCode (mirror for mainland China): https://gitcode.com/qq_35054471/virtual-world
• GitHub: https://github.com/miduo100/3d-virtual-world

Plan Comparison: Self-Hosted Deployment vs. Platform Leasing

About Genesis

Genesis is a self-hosted 3D virtual world system built on Three.js + WebGL, helping individuals and businesses build their own 3D spaces. Accessible directly from a browser, compatible with both PC and mobile, it supports multiplayer online, federated teleportation, a shop system, and Agent integration—where an AI can enter your world as an embodied character. Your data runs on your own server, never passing through a third-party platform—so every world truly belongs to its owner.

Want dealers to get up to speed without waiting for central training or a sample unit? Genesis (the Genesis Virtual World CRM System) is a Three.js 3D virtual world foundation you deploy on your own server — open a browser and you are in, the training flow is defined by you, the data is yours, and the world stays. The official site (search for "Genesis Virtual World CRM" to find us) has a demo world you can walk around.

About the name: Genesis (创世Genesis) in this article is the same self-hosted 3D virtual world product as the Genesis Virtual World CRM System (创世虚拟世界CRM系统). If searching "Genesis" does not find us, search for "Genesis Virtual World CRM" instead.