RobotGuide
Deliberate practice/Interview Questions

Practice 2 / 4

Interview Questions

By role, 24 questions, and what counts as a sufficient answer

The interviewer never asks whether you know something; they ask whether you have made decisions you were accountable for. Each question below states what counts as enough — answers without numbers, without trade-offs, and without failure stories are not enough.

24 questions in total. Say it yourself before opening any answer.

Algorithms · Perception and SLAM 4

You say you have done SLAM — what accuracy?

Enough: give ATE / RPE numbers + test venue + sensor models + a comparison baseline (e.g., 30% better than Gmapping). Not enough: only 'it worked well'.

How does your detection degrade when lighting changes?

Enough: name the failure scenes (backlight, night, reflections) and your compensation (data augmentation / multi-exposure / infrared). Not enough: 'the model is strong'.

How do you handle multi-solution poses of symmetric objects?

Enough: explain the ICP initialization source, multi-hypothesis verification, and confirming with contact forces before grasping. Not enough: 'use deep learning'.

Why this backbone and not a bigger one?

Enough: give the latency budget (30 Hz = 33 ms) and the measured frame rate, showing that a bigger model busts the budget. Not enough: 'better results'.

Algorithms · Control and Motion 4

How do you keep a 1 kHz loop jitter-free?

Enough: measured jitter curves (worst case + distribution), CPU isolation, memory preallocation, lock-free communication. Not enough: 'we used a real-time OS'.

How do you choose MPC's prediction horizon?

Enough: tie it to the control period, the system time constant, and the compute budget, with concrete numbers and the measured solve time.

Why impedance control instead of commanding force directly?

Enough: explain contact stiffness, stability, and environment uncertainty — force control goes unstable against stiff environments; impedance gives compliance.

How do you guarantee no falls during gait switching?

Enough: explain support-phase detection, the state machine, and the fallback on failure (drop to walk or sit down actively).

System Software 4

How would you configure ROS 2 QoS?

Enough: per topic — sensor streams Best effort + queue 1, control commands Reliable + queue 1, and why not everything can use the defaults.

How do you do zero-copy?

Enough: explain loaned messages, shared memory, and which layer the serialization cost sits in.

A node is stuttering — how do you locate it?

Enough: ros2 topic delay, tracing, perf, and layered instrumentation — not 'try restarting'.

How do you sync clocks across machines?

Enough: PTP / chrony, the error magnitude, the impact on fusion; explain that without sync, lidar and odometry do not line up.

Embedded and Electronics Control 4

What happens at 80% CAN bus load?

Enough: latency jitter grows, low-priority frames starve, and worst-case response time is computed by priority (not by average).

What bandwidth should the FOC current loop have?

Enough: tie it to switching frequency, sampling delay, and the motor's electrical time constant, with a concrete set of numbers.

The board powers up but does not boot — how do you debug it?

Enough: watch the serial console through BootROM → U-Boot → kernel → init, measuring power and reset first.

How do you detect encoder pulse loss?

Enough: ABZ plausibility checks, velocity-jump detection, dual-readout comparison — not 'it ran and looked fine'.

Mechanics and Structures 4

Joint stiffness is insufficient — what now?

Enough: first distinguish drivetrain backlash, structural compliance, and control bandwidth — the fixes are entirely different.

Benefits and costs of a large gear ratio?

Enough: reflected inertia falls by i² and torque multiplies by i; the costs are lower speed and a bigger backlash influence.

How do you handle heat?

Enough: give measured continuous torque and temperature rise, the heat path, and the derating strategy.

How do you guarantee assembly precision?

Enough: tolerance-chain analysis + calibration compensation + measured verification — all three, no exceptions.

Testing and Integration 4

How do you design a full-machine test?

Enough: layering (unit / bench / full machine / field) + pass criteria + a rollback plan on failure.

How do you reproduce an intermittent fault?

Enough: log environmental variables, run longer monitoring, amplify stress (heat / pressure / noise) — not waiting for it to appear on its own.

Which metrics do you check before launch?

Enough: success rate, MTBF, worst-case latency, fault recovery time — numbers for all four.

How do field and lab differ?

Enough: ground, lighting, network, temperature, human interference, with your countermeasure for each.