Onboard Compute Platform

The main computer mounted on a robot that runs perception and model inference.

The onboard compute platform (主控, sometimes just “main control” in Chinese industry usage) is the compute hardware carried on a robot's body that does the “thinking” — running perception, localization and navigation, task planning, and inference for models such as a VLA — and then sends commands down to the motion controller and each joint's driver. Common choices include NVIDIA's Jetson line (Orin, Thor), Rockchip's RK3588, NPU-equipped SoCs like D-Robotics' RDK S100, and x86 industrial PCs. Selection mainly comes down to compute throughput (measured in TOPS or TFLOPS), memory, power draw, available interfaces, and software ecosystem. Because a robot runs on battery and has to manage heat, onboard compute is far weaker than a training server's, so models are often compressed, quantized, or partly offloaded to the cloud. In Chinese, 主控 can also loosely mean “the main chip on a control board,” so the meaning depends on context.

ExampleNVIDIA's Jetson Thor module is rated at up to 2,070 FP4 TFLOPS with 128 GB of memory and a 40–130 W power range, positioned as the onboard compute platform for humanoids and other robots.

Also called

Compute Platform, Edge Compute Platform, Main Controller

Related

NVIDIA Jetson · NVIDIA Jetson Thor · D-Robotics RDK S100 · Rockchip RK3588 · Lower-Level Controller · On-Device / Edge Deployment

Sources

Jetson Thor | NVIDIA

As of

2026-09