Skip to content
Tech Interview Prep

Real Interview Questions for the Robotics Engineer at Tesla

New

Practice MCQs

Answers count toward the Robotics Engineer and overall leaderboards; your latest answer to each question is the one that counts.

21 questions · Flow follows Relevance until this set has a learning order.

MCQ-1An engineer calls a 12 ms lidar-to-odometry skew her hardest problem: 120 AMRs, a failed pick every ~4,000 cycles, 40,000 cycles of logs she could replay offline, and a CAN bus at 30% load. The skew never showed in any log, because the node that published the lidar also stamped the odometry with its own clock. The fix was one line and recovered 3% throughput, about $1.2M a year. Which account of why the problem was hard is right?
MCQ-2A candidate's grasp policy misses 30% of her lab's 200-object benchmark; every miss is an object wider than the gripper's fixed 20 mm stroke, and the widest object is 55 mm. A rigid 40 mm jaw bolted on dropped the same policy to 40%, since it had never seen that stroke. The gripper is a vendor part with no stroke option, the protocol is frozen, and the cell has one arm. Asked why Tesla's Optimus team over a traditional AI lab, which answer does her record support?
MCQ-3A humanoid's elbow joint has no reference design, so the team sizes it from the load case: with the forearm horizontal, a 5 kg payload sits as a point mass 0.30 m from the axis, the 1.8 kg forearm's center of mass sits 0.15 m from it, and the joint accelerates at 10 rad/s². The motor peaks at 0.8 N·m through a 30:1 gearbox at 85% efficiency. Which account of the joint's capability is right?
MCQ-4A dispense-height offset for the adhesive cell must reach the line before the 06:00 trial, the only downtime window this week. The full hardware-in-the-loop regression takes 6 hours; the rig is free only 04:00-08:00, when another team takes it. The change is one vision-to-motion constant, recomputed per part type at changeover; the safety controller binary is untouched. The trial runs one part type at 25% rate with an operator on the E-stop. What do you run on the rig, and what do you defer?
MCQ-5A Tesla humanoid team's manager wants the knee actuator's firmware torque limit raised from 40 N·m to 52 N·m for Friday's demo. Your bench log shows the gearbox case reaching 118 °C at 52 N·m sustained, against a 105 °C grease rating, while the demo's duty cycle peaks at 38 N·m. The manager asked for the bench data by Wednesday, and the safety sign-off owner chairs Thursday's review. What do you do?
MCQ-6Your humanoid wrist actuator shipped after a bench qual of 500 ten-minute cycles at 25 °C ambient, housing never above 55 °C, drift under 0.02°. Three of twelve field units now drift past 0.5°; each ran 40-minute continuous shifts at 40 °C ambient, logged housing above 85 °C for 30 minutes before drifting, and came from a different encoder lot. The 40-minute duty is a customer requirement. Which change does the evidence support?
MCQ-7Optimus must hand a 1.2 kg torque wrench to an operator at the inspection station by the end of Q3, nine weeks out. The arm's reach and 2 kg payload were qualified at this station last quarter; the gripper, the operator's gloved hand arriving anywhere within ±150 mm, and the release are new. Safety sign-off takes four weeks, and the line is free for trials only on Sundays. Which milestone plan fits?
MCQ-8A humanoid knee must hold 60 N·m peak and 25 N·m continuous over its duty cycle, and a technician must be able to backdrive the output by hand against no more than 3 N·m. The motor makes 7 N·m peak and 1.2 N·m continuous with 0.03 N·m detent, and the team models hand-backdrive torque as gear ratio × detent plus the gearbox's breakaway friction. Which gearbox clears all three limits?
MCQ-9A joint's velocity loop is tuned with a 10 kg payload (reflected inertia 0.020 kg·m²) to a 60 Hz crossover and 50° phase margin; the drivetrain resonates at 180 Hz. With the gripper empty (0.005 kg·m²) the joint buzzes at 180 Hz. The team wants one gain set that holds 60 Hz crossover and 50° margin for both payloads, switching on a measured inertia estimate; the filter chain is frozen and the motor has torque headroom. Which change fits?
MCQ-10A 120:1 strain-wave shoulder joint has run 4,000 h at 12 N·m. In the last 300 h, output drag at 30 rpm rose from 0.35 to 0.95 N·m and housing temperature from 71 to 94 °C, while lost motion moved only 0.010° to 0.012°, grease iron held at its 45 ppm background, and the shaft- and mesh-frequency vibration lines stayed flat. Those 300 h began with a service that refilled the cavity with a lithium-complex grease; the factory fill is polyurea. Which diagnosis and corrective action fit the log?
MCQ-11The elbow stator is packaged at 40 mm OD × 25 mm with a 30:1 gearbox chosen for the joint's top speed, and the payload spec allows no derate of the rated continuous torque. At that torque copper loss is 35 W; at 600 Hz the 0.35 mm laminations add 45 W of eddy loss, which scales with thickness squared. The 0.20 mm gauge raises copper loss 4% through its lower stacking factor. The sealed housing rejects 60 W at the joint's 40 °C ambient. Which build should the team specify?
MCQ-12A bin-pick policy scores 94% over 500 sim episodes and 61% over 200 cell trials on the same SKU list, and all 78 physical failures are objects over 350 g. The sim already runs the cell's 100 Hz loop, its measured 40 ms perception latency and friction values taken from those same parts; its gripper model commands 30 N with no ceiling, while the cell saturates at 22 N, the motor's continuous thermal rating. Which replay isolates the cause?
MCQ-13A 6-DOF arm's IK uses the undamped pseudoinverse of its geometric Jacobian. At a wrist singularity the smallest singular value is 0.00024 m/rad, the target is 0.5 mm away, and the solver commands a 2.1 rad step that drives the elbow past its 2.0 rad limit from 1.95 rad. The servo tick is 1 ms and the cell allots 20 ticks for the move. Which solver change fits this configuration?
MCQ-14A 100 Hz Kalman filter with a fixed constant-velocity model tracks parts on a conveyor from a 30 Hz camera whose 60 ms transport delay is already compensated by timestamp. At a steady 1.5 m/s the position innovation is zero-mean with 1.5 mm spread, matching the camera's static-target repeatability; through a 0.4 s ramp at 2 m/s² it holds a steady +9 mm mean with the same spread, and +4.5 mm at 1 m/s². Which change fits this evidence?
MCQ-15A bin-picking planner's broad phase flags a grasp as colliding, so the planner runs a separating-axis test on the pair. Box A has half-extents 0.10, 0.05, 0.05 m and is axis-aligned at the origin; box B has the same half-extents, is centred at (0.20, 0.15, 0) m and is rotated 45° about z. Which account of the pair is right?
MCQ-16Optimus must insert a connector into a port with 1.5 mm clearance, and the approach runs at the safety layer's 0.1 m/s Cartesian speed cap. The policy is trained on 200 teleop demos recorded at 30 Hz with Cartesian end-effector poses; onboard inference takes 55 ms and the impedance controller runs at 200 Hz. Re-running inference on the same observation returns end-effector targets that differ by up to 3 mm. Which output and execution scheme fits?
MCQ-17Two AMRs each run their own SLAM and both publish `map -> odom` at 50 Hz under identical frame names. A palletizing arm looks up `map -> base_link` for each robot at 20 Hz; over a 45-minute shift the two maps drift 0.35 m apart and the arm's target for robot 2 jumps by that much on each re-resolve. The robots meet only at a shared transfer station, and their radio link drops for minutes at a time. Which change fixes the lookup?
MCQ-18A 20 kHz FOC current loop on a 168 MHz Cortex-M4F has 4.0 µs of slack in its 50 µs tick, and the `sinf`/`cosf` pair in its Park transform costs 6.4 µs. The angle comes from a 17-bit absolute encoder (0.0027° per count), and the torque-ripple budget lets the trig approximation add at most a tenth of one count. Flash has 8 KB free and SRAM 1.5 KB. Which replacement fits?
MCQ-19A humanoid knee joint trips `MOTOR_OVERCURRENT` about once per 2,000 walking steps. The motor-controller firmware logs 48 A for 2 ms against a 45 A limit, but the joint's output torque sensor reads 30 N·m at that instant; with a 30:1 gearbox at 85% efficiency and a 0.05 N·m/A motor, that torque implies about 24 A. A bench run at a constant 30 N·m never trips, and the encoder and EtherCAT trajectory stream show no discontinuity. Which test isolates the layer?
MCQ-20A weld cell runs 400 robots, and each J2 RV reducer fails on average once per 5 robot-years. An unplanned seizure costs $200,000 in downtime and scrap; a planned swap costs $20,000 and needs 14 days' notice to fit a weekend window. The team's classifier outputs a calibrated probability that a reducer will fail within the next 7 days, but measurable degradation starts 21 days before seizure. Which model and decision rule fit these numbers?
MCQ-21A 48 V, 8-pole-pair BLDC joint motor returns after a logged 4 s stall with winding temperature at 120 °C. Holding 1.2 N·m at 25 °C now draws 14.4 A where the bench record shows 12.6 A; open-circuit line-to-line back-EMF at 1,000 rpm is 8.8 V against 10.0 V when new; phase resistances are 0.421, 0.419 and 0.423 Ω; 1× and BPFO vibration and bearing temperature are unchanged. Which diagnosis and corrective action fit the evidence?

Curated: · Written: · Reviewed: