Six self-test questions
Answer first, then expand the reference answer. Missing one is fine — knowing which block is missing is the point. These six also work as questions for you to ask the interviewer.
Q1
Jitter budget for a 1kHz control loop: period 1000µs — how much jitter does your system allow? What is eating it?
Show reference answer
Under PREEMPT_RT, 10–50µs is normal and holds up a 1kHz loop. Four things at the config level: isolate cores (isolcpus / nohz_full / rcu_nocbs), SCHED_FIFO priority, mlockall to lock memory, and disable deep C-states and frequency scaling. What eats jitter: memory paging, non-real-time drivers, interrupts stealing cores, CPU frequency scaling.
Missing it means: You only know the term 'real-time Linux' and have never run a controlled experiment. fill in Real-time systems & PREEMPT_RT →
Q2
10 joints need 1kHz synchronization. Can CAN do it? Why can EtherCAT?
Show reference answer
No. At CAN 1Mbps, one 8-byte frame takes about 130µs including arbitration and transmission; 10 axes queuing up is 1.3–1.6ms, so the practical CANopen sync cycle tops out around 500Hz. EtherCAT relies on slaves reading and writing on the fly (frames never stop) plus distributed clock sync, with jitter under 1µs and cycles down to 125–1000µs.
Missing it means: Treating the bus as 'just pick an expensive one', not knowing it's a hard ceiling. fill in Fieldbus networking & real-time behavior (EtherCAT / Ethernet / Modbus) →
Q3
In ROS 2, should a 1kHz control topic use RELIABLE or BEST_EFFORT QoS?
Show reference answer
BEST_EFFORT with queue depth 1. RELIABLE retransmits on loss, and retransmission means this cycle's data arrives late — latency becomes uncertain, worse than dropping a frame. Perception and point clouds can be RELIABLE. The criterion: does losing one frame make the system lose control.
Missing it means: You can use ROS 2 but never thought architecturally about 'which topic is allowed to drop'. fill in DDS & QoS →
Q4
Add a gear ratio — does the output get faster or slower?
Show reference answer
Slower. With ratio G added, the motor sees a lighter load (divided by G²) but the output sees a heavier motor inertia (multiplied by G²), and backlash and compliance are amplified. Harmonic joints resonate around 30–80Hz, capping force-control bandwidth at 10–20Hz.
Missing it means: Selecting actuators on the 'bigger torque is better' instinct, ending up with a joint you can't tune. fill in Mechanical design & transmission →
Q5
Torque-loop latency must be <= 1ms — how do you back-solve to motor inductance?
Show reference answer
Torque loop <= 1ms → current loop bandwidth >= 1kHz (2kHz for high-end force control) → sampling rate >= 10kHz (bandwidth is about 1/10–1/20 of the sampling rate) → τ = L/R <= 1ms → inductance L <= 0.2–0.6mH (48V motors often have R of 0.2–0.6Ω) → low inductance brings big ripple → PWM at 16–20kHz with double update → bus <= 250µs → must be EtherCAT → inertia ratio: direct drive <= 5:1, with gearbox <= 20–30:1 → encoder >= 17bit on the motor side, >= 19bit on the output side.
Missing it means: You don't yet think of parameters as mutually constraining; on datasheets you only compare magnitudes. fill in Motors & FOC drives →
Q6
Where does the line between host and low-level controller sit? What do Jetson and MCU each do?
Show reference answer
Anything with a cycle <= 1ms sinks to the MCU. Jetson runs ROS 2 perception, planning, and orchestration (10–100Hz); the MCU runs current and torque loops (1kHz and up), bridged via micro-ROS or CAN. The criterion isn't 'which chip is stronger' but the jitter requirement: even with PREEMPT_RT, Linux only has 10–50µs of headroom.
Missing it means: You can write ROS 2 nodes but never asked 'where should this node run' — which is precisely the first question of architecture interviews. fill in Robot software architecture →
How to read the score
| Correct | Meaning | Next step |
|---|---|---|
| 0–2 | The gap is real-time systems and domain knowledge, not coding | Do not apply yet. Fill systems and robotics fundamentals first |
| 3–4 | You can apply, but bring measured data | Turn questions 1 and 5 into a project with three compared runs |
| 5–6 | Apply directly | Aim the resume at the roles page; send 8–10 applications |