Sending operators onto a real excavator to practice pre-op inspection is expensive, ties up the machine, and carries real safety risk. The usual fallback, slide decks or video lectures, gets clicked through without the procedural knowledge actually sticking. I built a browser-based training simulator as a personal proof-of-concept to test a different approach: close to the real thing, without the cost of a full 3D asset pipeline.
The core idea: 3D for movement, AI images for detail
The simulator walks the learner through 14 checkpoints in a structured pre-op inspection. A virtual instructor, Dave, narrates each step by voice. Real-time 3D handles navigation and spatial orientation, while AI-generated images drop in during static camera phases to deliver photorealistic detail on specific components without the cost of building and rendering that detail in the engine.
That hybrid was deliberate. Real-time rendering struggles to match photographic detail on close-up inspection views without a heavy asset pipeline, but AI image generation could produce highly detailed views of hydraulic lines, fluid reservoirs, and safety decals to drop in exactly where the learner needed to actually look closely.
Fourteen checkpoints, following the real procedure
The training follows the same linear workflow as an actual site inspection: tracks, undercarriage, hydraulic hoses, engine oil, coolant, handrails, steps, fire extinguisher, and more. At each station Dave explains what to look for, what counts as a defect, and why it matters, with AI-generated variations showing realistic damage, corrosion, bent metal, missing bolts, without needing to photograph actual broken equipment.
Why voice, not text
I didn't want learners reading walls of text on a safety module. Dave narrates every instruction and delivers the knowledge checks aloud, in a consistent, professional tone close to what you'd expect from a site supervisor. In testing, users who kept audio on finished faster and scored higher on the final knowledge check than those relying on text alone.
Knowledge checks land at four points during the flow, right after the relevant checkpoint, not bundled at the end. A question about hydraulic hose condition appearing immediately after the learner studied the hose images and heard the explanation measurably improved retention over a single end-of-module test.
Built to run anywhere, including offline
The whole experience runs in the browser on vanilla JavaScript, HTML, and CSS, with state managed client-side so it keeps working offline after the first load. That constraint was intentional: many industrial sites have limited or unreliable connectivity, and a training tool that depends on a live connection isn't one operators can actually use on the floor.
What this proved
The inputs were reference images, publicly available machine documentation, and a browser. No proprietary assets, no game engine license, no photogrammetry scan of a real excavator. The output is a structured, voiced, interactive training experience that holds up both visually and procedurally, which says more about how far AI-generated imagery and TTS can now stretch a limited budget than it does about this specific machine.
If your organization needs job-specific training software, heavy equipment, factory machinery, safety procedures, pre-operational checks, this is the kind of tool I build: grounded in the real procedure, without the overhead of a full 3D production pipeline.