This year’s WAIC and WRC showed a shift in embodied intelligence: less showboating, more pragmatism. Last year’s robots danced, fought and walked runways; this year they were dropped into real logistics, industrial and home tasks. That points to a bigger question: how far has embodied intelligence actually come, and how far to real scale?
From a diffusion view, a technology moves from innovators and early adopters to the early majority. What decides whether it crosses into sustained growth is clearing the gap between early adopters and the early majority. Today embodied intelligence is still in the early-adopter phase. One insider’s fairly optimistic call is two to three years, but first come hard problems: real-world data from millions to tens of millions of hours, base models that are stable and efficient rather than merely capable, hardware cost and reliability at labour-replacement standard, and To B scenarios that truly clear return-on-investment.
Leiphone spent two months visiting embodied-AI firms and watching demos, and one feeling grew stronger: once a robot actually works, it needs far more than a good body, because real tasks are far messier than demo scripts. Crossing from showing ability to completing tasks sits behind a whole stack of systems engineering, and the cloud is becoming a critical piece of it.
Putting robots on the cloud
A Tencent Cloud expert argues, like autonomous driving and large models, embodied intelligence shows a long-tail effect: the first 60, 70 or 80 per cent may come fast, but the last 20 per cent takes disproportionately long. Filling that last 20 per cent needs a full systems stack. IoAI’s teleoperation business started local, with the operator beside the robot, needing no network. But remote scenarios demanded stable, safe control and image streams plus cloud compute to help under weak networks.
Tencent Cloud’s TRRO (Tencent Cloud Remote Real-Time Operation) took that need. Its experts note embodied systems are inherently heterogeneous: not every capability belongs on the body. Slow-reaction tasks like task and scene understanding go to the cloud on high-performance GPUs for training and complex inference; fast-reaction tasks like motion control and real-time response stay on the edge with small NPU or CPU models, cheaper and lower latency. That split is itself a way to cut deployment cost and lift adaptability.
Data is the hard bone
Once robots connect to the cloud, the next problem is data handling. Lingchu hit this early: as collection scaled, one data factory could produce hundreds of gigabytes a day, and storage, transfer and processing needs exploded. For a startup, building that alone is poor value, so partnering with a cloud provider fits. Tencent Cloud’s storage architect describes cooperation across three layers: data platform, training platform and audio-video streams.
Embodied data is more complex than plain images or video: each episode carries task, action and other structure alongside video, joint data, control commands and labels. Lingsheng is converting VLA to WAM paradigms and struggles with storage scale, fragmentation and retrieval, while wanting a two-way loop where data trains models and models help process data. The value of cloud shows most at the world-model pretraining stage, where data scale and quality demands are highest.
When the Agent starts scheduling the robot
Yuandian Robot and Tencent Cloud show what happens when an Agent runs on the robot. Its future home products split into Pet, Partner and Assistant, all scheduled by an Agent that understands user needs, generates skills and dispatches tasks. After connecting Tencent Cloud ClawPro, a request like having the robot dance when a stock rises becomes a skill the Agent schedules rather than a newly coded feature.
To keep an Agent running long term needs deeper runtime infrastructure. Agent Runtime rests on four layers: a sandbox for the environment, a bucket for personal data, memory for long-term recall, and a security system for permissions. Together they give robots a long-running, memory-bearing, safely resourced substrate, the exact foundation future home robots need.
As robots enter more scenarios, demand for data, compute, connectivity and Agents rises, and the cloud’s role stretches beyond traditional compute and storage. One analogy from the visits: robot deployment is a fist, and data, compute, model, scenario and body are the five fingers. Only when they close together is the fist real. Embodied commercialisation is fundamentally systems engineering, and the cloud is becoming its indispensable infrastructure layer.
Editor’s note: This is an adapted translation of the original Leiphone report. It has been trimmed and restructured for readability for an international business audience.
Translated and adapted from Leiphone (https://www.leiphone.com/category/industrycloud/31pSSHUoPdFy7U3E.html).