Rokid

Solution

Rokid refactored its spatial-computing pipeline around Alibaba Cloud Function Compute so the GPU-intensive 3D-mapping portion of the workflow could run as a true serverless service. The company packages its processing logic into container images and lets Function Compute allocate GPU resources on demand, which removes the manual process of requesting, operating, and releasing ECI instances for each job. Function Compute's asynchronous system also lets Rokid chain multi-step 3D-modeling tasks so the output of one stage automatically triggers the next instead of relying on a separate coordination service. Built-in metrics, alerts, traces, and logs improve observability across the whole processing lifecycle, while warm-resource capabilities and reserved instances reduce the latency of subsequent invocations. For the real-time scene-experience module, where sub-second response is mandatory, Rokid still keeps ECI capacity scheduled ahead of peak times, giving the company a split architecture optimized for both offline GPU throughput and real-time AR response.

Data flow

Captured venue video and scene assets are split into asynchronous processing stages and sent to Function Compute GPU jobs launched from container images. Each stage emits outputs that trigger the next step in the 3D-modeling pipeline, while the real-time AR positioning service is served separately from scheduled ECI capacity.