Inside a coal mine hundreds of kilometres from any game studio, an operator in an AR headset swings a virtual excavator arm while sitting in an air-conditioned room. The software steering that machine was born not in a factory but in NetEase Fuxi (网易伏羲), the gaming company's AI lab, where the first customers were non-player characters. The unlikely export of "game tech" into heavy industry is becoming one of China's more original corporate AI stories.
Key takeaways
- NetEase Fuxi (网易伏羲), founded in 2017, pioneered in-game AI agents such as intelligent NPCs in Justice Mobile (逆水寒) and voice AI teammates in Naraka: Bladepoint Mobile (永劫无间).
- Its UGC "film-crew mode" (剧组模式) let players generate animations from text; one title saw over 5 million player-made clips within two weeks of launch.
- The NetEase Lingdong (网易灵动) robot unit applies game-derived teleoperation and digital-twin tech to excavators and loaders, with one operator managing multiple machines.
- A single Lingdong excavator is reported to move over 5,000 cubic metres per day, with efficiency gains of 90–120% versus traditional sites.
- The playbook — build AI in games, then port it to physical industry — is a differentiated, low-hype path worth global attention.
Game AI as a training ground
Fuxi was China's first game-artificial-intelligence laboratory, set up inside NetEase in 2017. Rather than chasing general models, it optimized for interaction: characters that remember, teammates that take voice commands, and player tools that turn a sentence into an animation. In Justice Mobile (逆水寒), LLM-driven NPCs gained memory and multi-modal behaviour; in Naraka: Bladepoint Mobile (永劫无间), a voice AI teammate can "hear commands, chat, and fight," combining speech recognition, language models and reinforcement learning.
The economic logic is clever. Games are a safe, high-volume sandbox where AI behaviour can be tested on millions of players before any industrial risk. When Fuxi ported its text-to-animation "film-crew mode" (剧组模式) to Justice Mobile, players produced more than 5 million short clips in two weeks — a massive, free reinforcement signal no factory could supply. The result is a feedback loop unusual in heavy industry: consumer entertainment directly funds the research that later lowers the cost of autonomous machinery, blurring the line between a game studio and an industrial-software supplier.
From pixels to excavators
The leap to physical industry came through NetEase Lingdong (网易灵动), Fuxi's robotics brand launched at WAIC 2024. Its core trick is borrowed straight from game engines: low-latency audio-video transmission (end-to-end under 90 ms) and physics-based digital twins let a remote operator "play" heavy machinery in first person. In Inner Mongolia open-pit mines, Lingdong excavators run 24-hour unmanned shifts moving over 5,000 cubic metres daily — roughly the work of 300 skilled operators. The "shepherd" (牧羊人) mode lets one operator supervise several robots at once.
In November 2024, NetEase signed a strategic deal with XCMG (徐工), a major construction-equipment maker, to co-build an unmanned "black-light" (黑灯) batching-plant product and the multi-machine mining mode. Reported outcomes: productivity at 120% of traditional sites, energy use down 15%, and overall operating cost down 30%.
Local media reported that Lingdong's initial model-training investment alone reached tens of millions of RMB (on the order of $10+ million / HK$80+ million) over three-plus years — modest next to frontier-lab budgets, which is partly why NetEase targets two narrow products rather than horizontal infrastructure.
Why this travels
Most Chinese AI narratives are about bigger models or more compute. Fuxi's is about reuse: decades of game-rendering, simulation and interaction research become the control layer for real machines. The approach is already spreading beyond mines to ports and emergency response, and NetEase says it is exploring "Belt and Road" exports of its smart-mining stack. The broader significance is that Fuxi treats entertainment software as a low-risk laboratory for industrial control systems. By validating behaviour in a virtual world first, the lab converts playful interaction research into operational reliability before any physical machine moves, inverting the usual order in which industrial automation is built from scratch and only later made usable.
For global readers, the lesson is that valuable AI is not only born in labs — it can be harvested from entertainment software and dropped into sectors desperate for automation. As NetEase CEO Ding Lei has put it publicly, the company is building a "virtual-to-real" technology loop in which games are both a starting point and a hub connecting industries.
Honest limitations
Lingdong's efficiency and cost figures are company- and partner-reported, drawn from specific mine and batching-plant deployments rather than large randomized studies. The "300 operators" equivalence and 5,000-cubic-metre daily number describe highlighted sites, not a fleet average. Fuxi's game-AI leadership is real but narrow; it is not a contender in general-purpose frontier models. Deployment scale — "dozens of provinces, 50-plus projects" — is promotional and should be verified per site before investment.
What readers can do now
- Track NetEase Lingdong (网易灵动) as a case study in cross-industry AI reuse, not as a model vendor.
- For mining or port automation, request site-specific references rather than relying on aggregate efficiency claims.
- Watch WAIC 2026 and XCMG partnership updates for evidence the "shepherd" mode scales beyond pilot mines.