Systems & software · topic D8 · 40 hours
Robot software architecture
The architecture track interviews exactly this: who can organize dozens of modules, multiple frequencies, and multiple real-time levels into a system that can evolve for ten years without falling apart.
What to learn
- Layering: the frequency and real-time level of each of perception / decision / control / actuation, and where to draw the boundaries
- Module boundaries and interface contracts, dependency governance (who may call whom, and how to prevent chaotic calls)
- Failure isolation and degradation strategy: when one node dies, how the machine as a whole should survive
- Split of duties between the high-level board and the low-level board, OTA and version management, layering of configuration and parameters
The lesson · plain words
Software architecture is not drawing block diagrams; it is deciding «who may be late, and who cannot be late by even one millisecond». Layering in robot software has one core principle: physically separate the hard real-time parts from the compute-hungry parts.
This page drives that principle home with one frequency table.
The frequency pyramid: the lower, the faster
Watch the figure: the sensing/drive/control layers start at 1 kHz with latency budgets of a few tenths of a millisecond; planning runs at 100 Hz; perception runs at tens of Hz with tens of milliseconds per frame. Any architecture that pushes perception into the control loop makes the 1 kHz servo loop wait for a 20 ms detection — a disaster.
The right move is «separate fast from slow»: the real-time layer (RTOS or RT kernel) does only sampling–control–output; the Linux layer does perception and planning; the two connect asynchronously via shared memory/DDS, so slower perception never disturbs the servo loop's beat.
Interactive figure · drag it and see
Three typical architectures
① Monolithic real-time: everything inside one MCU — simple and controllable, low compute ceiling — lab rovers and drone flight controllers. ② Layered hybrid: an RTOS control layer + a Linux intelligence layer — the mainstream product form (robot-arm controllers, AGVs). ③ Cloud–edge collaboration: the edge does real time, the cloud does compute — service robots and fleet scheduling.
One-line selection rule: first count how long your hard-real-time checklist is, then decide whether you need a second computer. The classic team mistake is going straight to an «all-Linux stack» and then paying the debt back in real-time performance.
Interface design: what the layers say to each other
Between the real-time layer and the intelligence layer only «state» and «commands» cross — never raw data; raw data (images, point clouds) takes a separate bypass. This one discipline guarantees that when the intelligence layer dies, the real-time layer can still shut the machine down safely.
Define state and commands as versioned structures (adding fields must not break old parsers); make error semantics explicit (a STATUS field, not guessed from timeouts). Interfaces are the skeleton of the architecture — changing an interface costs ten times more than changing an algorithm.
Evolution: architecture is iterated into shape
No architecture is right on the first draft. A workable path: get the monolith running first → split the slow parts into separate processes → move the real-time parts into an RT kernel/RTOS → accept each layer against the frequency pyramid. Every step has standalone value; do not try to leap to the summit in one go.
Acceptance method: one frequency report per process (actual Hz vs expected) and one resource report (CPU/memory). Pin both reports in the repo README, and the health of the architecture is visible at a glance.
Three questions — do you really get it?
1. You put a 30 Hz vision detection into a 1 kHz control loop and wait synchronously — the consequence?
- AMore accurate detection
- BThe control loop slows by two orders of magnitude and the beat falls apart
- CNo impact
Show answer
B. Fast and slow must be separated: perception delivers results asynchronously, and the control loop never waits.
2. What should pass between the real-time layer and the intelligence layer?
- ARaw images and point clouds
- BState and commands, with raw data on a bypass
- CEverything
Show answer
B. Minimal interfaces + a data bypass guarantee the real-time layer can still catch the machine safely when the intelligence layer crashes.
3. The core motivation of the layered hybrid architecture (RTOS + Linux) is?
- APower savings
- BPhysically isolating hard real-time from compute-hungry tasks so each gets what it needs
- CEasier coding
Show answer
B. Both sides matter: the servo loop needs deterministic latency, vision needs massive compute. One machine, two worlds, joined at the interface.
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
- A beautiful block diagram with no answer for each layer's latency budget or failure modes
- No degradation path: perception crashes and the whole machine stops; on site the only move left is a reboot
You have learned enough when
The Chinese edition carries the full topic page for Robot software architecture, including the complete outline, depth requirements per career track, and further derivations: robotguide.dev/k/D8/.