A delivery robot starts from a known position and navigates to a destination using only wheel encoders (dead reckoning). After 500 m, its estimated position is 8 m off. What is the root cause of this growing error?

The encoder batteries run low, reducing measurement frequency Small wheel slip, encoder quantisation error, and uneven terrain cause tiny positional errors at each step, which are integrated (summed) over the full journey so the error grows continuously with distance The robot's map is inaccurate so it cannot navigate correctly Dead reckoning only works for straight-line travel, not turns

A SLAM algorithm is running on an indoor robot. After exploring a new corridor, the robot re-enters a room it visited 20 minutes earlier and recognises distinctive features from its earlier scan. The algorithm uses this recognition event to dramatically reduce position error. What is this event called?

Dead reckoning reset Loop closure Occupancy grid update Path re-planning

A path planning algorithm uses A*. It finds the shortest path from a robot's current position to its goal through a large grid map. What is the key feature that makes A* faster than Dijkstra's algorithm when the goal location is known in advance?

A* does not check all possible paths — it uses random sampling to guess the shortest route A* runs multiple searches in parallel on different processor cores A* uses a heuristic function (estimated remaining cost to the goal) to prioritise exploring nodes that are likely on the optimal path, rather than exploring all nodes equally A* only works on graphs with fewer than 1,000 nodes

An engineer is planning the motion of a 7-DOF robot arm that must reach a cup on a table while avoiding a shelf above. Why is A* grid search unsuitable for this problem, and what algorithm is more appropriate?

A* is too slow for real-time use; potential fields are used for robot arm planning instead A* requires the robot to know the goal exactly; robot arms never know where the cup is A* requires a discretised grid; in 7-DOF joint space the grid has 7 dimensions making it computationally intractable. RRT is better because it samples the configuration space randomly without requiring a full grid A* cannot handle obstacles; only RRT guarantees obstacle-free paths

A robot using potential field navigation gets permanently stuck between two obstacles, even though a clear path to the goal exists. What fundamental limitation of the potential field method causes this?

Oscillation — the robot bounces between obstacles because the repulsive force is too strong Local minima — the repulsive and attractive forces can sum to zero at a non-goal location, creating a false equilibrium the robot cannot escape through gradient descent Path inflation — obstacles appear larger than they are, blocking the narrow passage Map resolution — the occupancy grid cells are too large to represent the gap between obstacles

In imitation learning (behavioural cloning), a robot vacuum is trained on 10,000 examples of a human manually guiding it around furniture. During deployment, it encounters a chair configuration never seen in training and fails. What fundamental limitation of behavioural cloning is revealed?

Overfitting — the robot memorised training examples and cannot generalise Underfitting — 10,000 examples are too few to train any useful policy Distribution shift — the robot only learned patterns seen in the training data and has no mechanism to generalise to novel situations outside the training distribution Reward hacking — the robot found a shortcut that works in training but fails in deployment

A reinforcement learning agent is training a simulated robot arm to reach a target. The reward function gives +1 only when the robot's gripper contacts the target. After millions of simulation steps, the robot still rarely receives any reward. What is this training problem called?

Reward hacking — the robot found a way to get +1 reward without reaching the target Sparse reward — the agent almost never receives a signal to reinforce good behaviour, making learning extremely slow or impossible without reward shaping Catastrophic forgetting — the robot forgets previously learned skills when training continues Overfitting — the reward signal is too specific to the training environment

A self-driving car's perception system correctly identifies every object in front of it. However, it cannot safely plan a path because it does not know what other road users will do next. Which module of the self-driving pipeline addresses this gap?

Perception — the perception module must also predict object motion Control — the controller adjusts speed based on proximity to other objects Prediction — forecasts the future trajectories and intentions of other road users before planning the ego-vehicle's own path Mapping — an updated HD map shows where other vehicles are allowed to go

An occupancy grid map represents a robot's environment as a grid of cells, each storing the probability of being occupied. A robot's LIDAR scans a corridor and the sensor model assigns probability 0.9 to cells where the beam ended and 0.3 to cells the beam passed through. Why use probabilistic values rather than simple binary (occupied/free)?

Binary maps require more memory storage than probability maps Sensors are noisy — probabilistic values allow the map to accumulate evidence from multiple scans and represent uncertainty, avoiding false certainty from a single noisy measurement Probability values allow the robot to move through partially occupied cells Binary maps cannot represent moving obstacles, but probability maps can

A self-driving car encounters a scenario never seen in its training data: a mattress has fallen off a truck and is lying flat on the motorway. Why is this 'long tail' scenario particularly dangerous for machine learning-based perception systems?

The car's sensors cannot physically detect flat objects lying on the road ML models trained primarily on common scenarios may misclassify or fail to detect rare objects, and their confidence scores may be unreliable for out-of-distribution inputs Self-driving cars require internet connectivity to look up rare objects, which may be unavailable The car's HD map does not include information about fallen cargo

A landmark-based localisation system places QR code markers at known positions throughout a warehouse. A robot with a camera detects a QR code and identifies its grid coordinates. How does the robot use this information to correct its estimated position?

The robot ignores its current estimate and jumps directly to the QR code's grid position The robot computes its position relative to the QR code using the camera image (distance and angle to the marker), then uses the marker's known global coordinates to calculate and update its own global position estimate The QR code broadcasts a GPS signal that the robot receives to determine its position The robot scans all QR codes in the warehouse simultaneously to triangulate its position

Waymo's self-driving cars have accumulated tens of millions of real-world miles, yet still require human oversight in new cities. What does this reveal about the relationship between training data volume and achieving full autonomy?

Waymo's systems are deliberately limited by regulation, not by technical capability Self-driving cars cannot handle any scenario not explicitly programmed — machine learning is insufficient Real-world driving has effectively unbounded complexity — more miles reduce but cannot eliminate rare dangerous scenarios, and new environments introduce new distributions the system has not yet learned The systems are fully autonomous but require human oversight for insurance and liability reasons only