RobotGuide

Hardware & electrical · topic E8 · 25 hours

Compute & domain-controller hardware

Robots today almost all use a layered compute structure of «domain controller + MCU»: the domain controller runs perception and planning, the MCU runs high-frequency control. A wrong pick (compute, interfaces, power, cooling) costs a huge rework bill — this is required reading for both electronics-control and system roles.

Core stage · Intermediate
Algorithm · Aware System software · Hands-on Embedded & controls · Proficient Mechanical · Aware

Tools
Tools

What to learn

  • Compute layering: who does what among MCU (current loops / above 1 kHz), SoC domain controller (perception and planning, e.g. Jetson Orin), and co-processors
  • Interface inventory: PCIe / USB / Ethernet / CAN / UART / GPIO — judge whether a board «looks like your system»
  • Power and thermal budget: TDP, passive vs active cooling, how to split the robot's total supply
  • Reading a board from the schematic side: power tree, boot-mode straps, debug ports (UART / JTAG / SWD)

The lesson · plain words

Main controller and domain controller selection is the robot's brain pick: the MCU handles real time, the SoC handles intelligence, the domain controller integrates both. Selection is fundamentally a multiplication problem: workload list × headroom × price.

The compute ledger: TOPS is not everything

Watch the figure: list the workloads one by one (model 12, planning 3, system 2…), then compare against platform compute. Note TOPS is a peak theoretical value; 40–60% is realistically usable (memory-bound, unsupported operators) — do the ledger in effective compute, not advertising compute.

Check the second dimension too: bandwidth. Large-model inference is often memory-bound, not compute-bound; without memory bandwidth, TOPS is wasted. Both columns belong on the selection sheet.

Interactive figure · drag it and see

What this figure shows: workload bars for three cameras, lidar, IMU/encoders, the detection model, planning, and system overhead, summed into a total demand that you compare against a draggable platform-compute (TOPS) slider, with a live verdict of overloaded / on the edge / headroom left. Takeaway: sizing a domain controller is three steps — list the workloads, add 30% headroom, match a price tier — and budget in effective compute, not peak TOPS.

Division of labor across the three hardware tiers

MCU (STM32/ESP32): microsecond-level real time, owns servo loops and safety. SoC (Raspberry Pi/Jetson/industrial PC): runs Linux + perception + planning. Domain controller (Orin/custom board): multi-camera input + large models + dual real-time systems — the standard humanoid-robot shape.

The optimum for small and mid projects is usually «one Jetson Orin Nano + one STM32»: the SoC handles intelligence, the MCU handles real time, bridged over UART/CAN. Do not charge into a domain controller on day one — its complexity will eat all your time.

The interface budget matters just as much

Cameras: MIPI CSI lane count and bandwidth (the multiplication problem of E19). Real-time buses: CAN-FD/EtherCAT controllers and channel count. Storage: eMMC/UFS read/write speed affects boot and data logging. Every interface is a scarce resource — list them before choosing a board.

Power and cooling are the domain controller's hidden budget: Orin draws 40 W at full speed, which passive cooling cannot hold — heatsink volume and airflow must be drawn into the mechanical layout at selection time.

Acceptance and expandability

Three acceptance tests: CPU/GPU utilization and temperature rise at full load (the flame graph of D10), measured bandwidth with all sensors sampling concurrently, and safe shutdown time on power loss. Only when all three pass does the platform stand.

Reserve expandability: 30% spare compute, one free CSI/USB3 port, one free CAN — the cost of attaching next-generation sensors decides how fast the product iterates.

Three questions — do you really get it?

1. When budgeting compute in TOPS, what should you do?

  • ACompare workloads directly against the nominal TOPS
  • BLook at the GPU only
  • CConvert at 40–60% effective utilization, then compare
Show answer

C. Peak compute is never reachable. Memory access, operator support, and scheduling overhead all discount it.

2. The pragmatic hardware combo for a small/mid robot project is?

  • AGo straight to a flagship domain controller
  • BOne SoC (e.g. Orin Nano) + one MCU real-time board
  • CMCUs everywhere
Show answer

B. Intelligence and real time each get a fitting home, bridging is simple, and cost stays under control.

3. Besides compute, what else must you check when picking a domain controller?

  • AMemory bandwidth and the interface inventory
  • BBrand prestige
  • COS version
Show answer

A. The memory bottleneck and scarce interfaces (CSI/CAN/storage) decide success just as much.

Keep going deeper

Links without a marker stay on this site; the rest point to the Chinese edition until its translation batch lands.

How it fails in the field

  • Picking a board by TOPS alone while ignoring interface count and real-time needs — perception runs, but control has nowhere to plug in
  • No thermal budget, so a summer field demo at full load throttles within an hour

You have learned enough when

Given a concrete requirement (how many joints, how many cameras, what control rate), produce a compute-layering plan plus an interface inventory, and explain why each interface must be reserved.

The Chinese edition carries the full topic page for Compute & domain-controller hardware, including the complete outline, depth requirements per career track, and further derivations: robotguide.dev/k/E8/.