AI Fitness Mirror based on Rockchip RK3588

Client

A home fitness startup aiming to provide real-time, AI-driven form correction and rep counting via a full-length smart mirror.

Challenge

Running complex 3D skeletal tracking algorithms at 30fps while simultaneously streaming 4K instructor video and rendering the UI, requiring massive compute bandwidth.

Product Concept & Architecture

The flagship Rockchip RK3588 (8-core CPU, Mali-G610 GPU, 6 TOPS NPU) is the absolute core of this system. It utilizes 8GB of LPDDR4x to handle the heavy memory bandwidth required by concurrent 4K decoding and AI inference.

Engineering Solution

We built a sophisticated software pipeline using Android 12. The 4K instructor stream is decoded via hardware VPU. Concurrently, a wide-angle 8MP camera feed is routed directly to the NPU using Rockchip's zero-copy memory architecture. The NPU runs a proprietary 33-point skeletal tracking model. The GPU then overlays the skeleton graphics onto the mirror UI in real-time.

Result

Zero perceptible lag between user movement and on-screen feedback. Sustained 6 TOPS NPU load with SoC temperatures remaining under 60°C due to a custom extruded aluminum heatsink.