Robotics Engineer interview questions, with worked answers

How to use these questions

Many robotics interviews start with a sketch: an arm, a target, an obstacle, a mobile robot in a corridor. Reason aloud from the geometry, the limits of the robot and the sensors, then say what you would test first on the real machine. Have one robot story ready: what it did, what broke, and how you found out.

What a robotics engineer is expected to know

• Motion: kinematics, motion planningPlan paths and trajectories that reach the goal, avoid collisions and respect the robot's limits. and control engineeringTune control loops on joints and drives so motion is fast, stable and precise..
• Senses: computer visionDetect and locate objects in camera images well enough for a robot to grasp or avoid them., sensor fusionCombine cameras, lidar, encoders and inertial sensors into one estimate the robot can trust. and localisation and mappingLet a mobile robot build a map and know where it is on it, and spot when it is lost..
• Software: C++Write fast, safe C++ for the parts of the robot that must react in time., PythonScript tests, tools and data analysis, and prototype robot behaviour quickly. and ROSBuild robot software from ROS nodes and messages, and find a fault across them with the ROS tools., or the robot makers' own tools.
• The cell: PLCsConnect the robot to the line's PLC and read the signals and logic that run the cell., functional safetyDesign safety functions, such as stops and speed limits, to the level the risk asks for., commissioning and the people around the roboticsDesign, program and bring into service robot systems, from the arm or vehicle to the cell around it. system.

Tools come up side by side and none is a must: ROS 2 as a framework, MoveIt for arm motion, Gazebo, Isaac Sim or Webots for simulationTest motion and software in a simulator before the robot moves, and know its limits., MATLAB for modelling and control design. Panels want the reasoning more than the tool.

Motion and control

What is the difference between forward and inverse kinematics, and where does inverse kinematics get hard?

What they check: the geometry behind every robot move.

Short answer: "Kinematics describes how the joint angles of a robot set the position of its tool. Forward kinematics goes from joints to tool and has one answer. Inverse kinematics works backwards, from a wanted tool position to joint angles, and may have several answers, none, or awkward ones near a singularity, where the robot loses a direction of motion."

The full answer: Most strange robot motion near the edge of its reach is a kinematics problem, not a software bug.

A weak answer: "The robot software does that for you." Follow-up: "The arm flips its wrist halfway through a move." Say what is happening.

How does motion planning work for a robot arm, and how do you avoid a collision?

What they check: planning in a crowded space.

Short answer: "The planner searches for a path through the space of joint positions that reaches the goal without touching anything in the model of the cell, often with sampling methods. Collisions are avoided only as well as the model is right, so I keep the cell model up to date, add safety margins and check the first moves at low speed."

The full answer: The planner is only as good as the world you gave it.

A weak answer: "The planner finds a path, so it is safe." Follow-up: "Someone moved a table into the cell." Say what protects it.

How do you turn a path into a trajectory the robot can follow?

What they check: time, speed and limits.

Short answer: "A path is only the shape; the trajectory adds time: speeds and accelerations along the path that stay within the limits of each joint and of the load. I smooth the path, then time it so the robot moves as fast as allowed without jerks that shake the part or wear the gears."

The full answer: A smooth trajectory is often faster in the end than a sharp one.

A weak answer: "Full speed everywhere." Follow-up: "The gripper drops parts at the corners." Say what you change.

How do you tune a control loop on a robot joint?

What they check: practical control, not only formulas.

Short answer: "I model or measure the joint first, start with a cautious setting, then raise the gains while watching the step response for overshoot and noise, with the real load attached. I check stability margins, and I test at the edges: full payload, top speed and cold start."

The full answer: A loop tuned without the real load is tuned for a different robot.

A weak answer: "I copy the gains from the last project." Follow-up: "The joint oscillates only when the arm is stretched out." Say why.

Senses, software and simulation

Which sensors would you choose for a mobile robot, and how do you combine them?

What they check: sensors chosen for the job and their weaknesses.

Short answer: "Wheel encoders and an inertial sensor for motion, lidar or cameras for the surroundings, bumpers and safety scanners for people. Each fails differently: wheels slip, cameras hate glare, lidar sees through glass. Sensor fusion, for example with a Kalman filter, weighs them so one sensor's bad moment does not fool the robot."

The full answer: The question is always how each sensor fails, not how well it works.

A weak answer: "More sensors are always better." Follow-up: "The robot gets lost in a long, empty corridor." Say why.

How do you calibrate a camera to a robot, and how do you know it worked?

What they check: accuracy you can prove.

Short answer: "Calibration means measuring exactly where the camera sits relative to the robot and how its lens bends the image. I move a known target to many poses, solve for the transform, then check it: the robot touches points the camera finds, and I measure the error. I recheck after any knock or change."

The full answer: A calibration nobody checked is a guess with numbers.

A weak answer: "We calibrated it once at the start." Follow-up: "Picks drift off by a few millimetres over a week." Say what you suspect.

How does a mobile robot know where it is? Walk us through localisation and mapping.

What they check: the core of navigation.

Short answer: "The robot builds a map from its sensors and, while it moves, matches what it sees against the map to estimate its position; done together, this is simultaneous localisation and mapping. In production I usually map once, then localise against that map, and watch how sure the robot is, so it can stop safely when it gets lost."

The full answer: Warehouses change every day; a map is only right on the day it was made.

A weak answer: "GPS tells it where it is." Follow-up: "The shelves were rearranged overnight." Say what happens.

How would you build the perception for a robot that picks mixed parts from a bin?

What they check: perception that survives the real world.

Short answer: "A 3D camera over the bin, a model that finds parts and their grasp points, and rules for what to do when it is unsure. I test with real parts, real light and badly sorted bins, track the pick rate and every failure, and feed the hard cases back into training."

The full answer: The demo bin is tidy; the production bin never is.

A weak answer: "A trained model will recognise everything." Follow-up: "Shiny parts are missed in the afternoon." Say what you change.

How do you structure a robot's software in ROS, and what do you watch for?

What they check: robot software that can be tested and maintained.

Short answer: "I split the system into nodes with one job each, such as drivers, perception, planning and the task logic, connected by clear messages. I watch timing, because late messages are as bad as wrong ones, record data for replay, and test nodes apart before the whole robot."

The full answer: A robot you cannot replay is a robot you cannot debug.

A weak answer: "One big program is simpler." Follow-up: "The robot stutters only when the camera runs." Say where you look.

How do you use simulation, and how do you close the sim-to-real gap?

What they check: honest use of simulation.

Short answer: "Simulation lets me test motion, reach, cycles and software safely and early. Sim-to-real means the gap between how the robot behaves in the simulator and in reality: friction, sensor noise, delays and light that the model lacks. I add noise and variation in simulation and always check the key results on the real robot."

The full answer: Simulation answers "can it work?", the robot answers "does it?".

A weak answer: "It worked in simulation, so it will work." Follow-up: "The grasp that never fails in simulation fails one time in ten." Say why.

Safety, the cell and the senior round

Walk us through the risk assessment for a new robot cell.

What they check: safety as engineering.

Short answer: "A risk assessment lists every way the cell could hurt someone, in every phase from setting up to cleaning, rates each risk, and sets the measures that bring it down: fences, light curtains, speed limits, safe stops. I work through it with the safety team to the international standards for robots and machine controls, ISO 10218, ISO/TS 15066 and ISO 13849, and test every safety function before hand-over."

The full answer: The people who clean and repair the cell meet the robot in ways the designers forget.

A weak answer: "We put a fence around it." Follow-up: "Operators open the door to clear jams while it runs." Say what you change.

When can a collaborative robot work next to people without a fence?

What they check: knowing what collaborative really means.

Short answer: "A collaborative robot is built to share space with people, with limited force and speed or sensors that slow it down. But it is the application, not the robot, that is safe or not: a sharp tool on a gentle robot is still dangerous. The risk assessment decides, and I measure contact forces where the standard asks for it."

The full answer: "Collaborative" on the data sheet is a start, not an approval.

A weak answer: "Cobots are always safe." Follow-up: "The cell handles knives." Say what that changes.

The cell misses its cycle time. How do you find the lost seconds?

What they check: method under production pressure.

Short answer: "Cycle time is how long one part takes through the cell. I record a timeline of every step, robot moves, gripper actions, camera, conveyor and signals from the line, and look for waiting. Often the robot waits for something else, so overlapping steps or shorter moves win more than raising speed."

The full answer: The robot is usually the fastest thing in a slow cell.

A weak answer: "Turn up the speed." Follow-up: "Raising speed made the parts slip." Say what else you try.

What happens during commissioning, and how does the robot talk to the PLC?

What they check: getting a cell running on site.

Short answer: "Commissioning is bringing the cell into service: installing, calibrating, teaching positions, testing every signal and safety function, then running parts until it is stable. The robot exchanges signals with the line's PLC, such as start, ready, part present and fault, over a fieldbus; I agree that list with the controls engineer before arriving on site."

The full answer: The signal list agreed in the office saves the night shift on site.

A weak answer: "We sort the signals out on site." Follow-up: "The line stops whenever the robot faults." Say what you change.

As a senior robotics engineer or lead robotics engineer, how do you guide a junior robotics engineer through a first project?

What they check: growing people while delivering.

Short answer: "I give the junior engineer one part of the cell to own, such as the gripper and its sensors, with clear checks, review their design early, and take them to commissioning so they see the real problems. I explain safety decisions rather than just making them, and let them lead the next similar cell with me in the background."

The full answer: The best training in robotics is a real cell with a patient senior nearby.

A weak answer: "They learn by watching." Follow-up: "The junior engineer's change caused a crash in the cell." Say how you handle it.

Which parts of robotics work can AI take on today, and which decisions stay with the engineer?

What they check: sober use of new tools.

Short answer: "Learned models already help with perception and grasping, and AI assistants help write code, tests and documentation. I would still check every safety-related change by hand, never let a learned model bypass the safety functions, and keep the sign-off of a cell with engineers who answer for it."

The full answer: A robot does what its software says, including the confident mistakes.

A weak answer: "AI will program the robots." Follow-up: "A generated program moves the arm outside its safe zone in simulation." Say what you do.

Your CV for these questions

Each robotics line on your CV invites the questions above: which robot, what was hard, how you made it safe. Write each one so a lead engineer asks for the story behind it:

Hold your CV against a robotics engineer job with the free score; then prepare the robot story for the area it finds thin, from the sketch to the cell in production.