Sensors and Signal Processing
A robot that cannot sense its environment cannot act intelligently within it. Sensing is the first step in the sense-plan-act loop that underlies nearly all robotic behavior. But sensing is harder than it looks: the physical world does not hand a robot clean, precise numbers. It delivers a continuous stream of photons, pressure waves, magnetic fields, and photonic returns — all corrupted by noise, quantized by finite-resolution electronics, and filtered through physical optics or antenna geometry. Understanding how sensors work and how to process their output is fundamental to robot engineering.
A Taxonomy of Robot Sensors
Robot sensors are classified by what they measure and by whether they measure properties of the robot itself (proprioceptive) or properties of the external environment (exteroceptive). Proprioceptive sensors measure internal state. A wheel encoder uses a slotted disk and optical detector to count shaft rotations, yielding wheel velocity and, by integration, position. An inertial measurement unit (IMU) combines accelerometers (measuring linear acceleration) and gyroscopes (measuring angular velocity) to track the robot's own motion. Joint torque sensors in robot arms measure the force the robot is exerting — critical for safe human-robot interaction. Exteroceptive sensors measure the environment. Cameras capture a 2-D projection of the 3-D world; depth cameras (structured light or time-of-flight) add a distance estimate per pixel. Lidar (light detection and ranging) emits laser pulses and measures their round-trip time to build dense point clouds of the environment — the backbone of most autonomous vehicle perception systems. Ultrasonic sensors use sound waves for short-range proximity detection. GPS receivers triangulate global position from satellite signals but are unreliable indoors and in urban canyons. Each sensor type has a characteristic operating range, resolution, latency, power draw, and failure mode. Real robots typically fuse multiple sensor types, exploiting the strengths of each to cover the weaknesses of the others.
A camera provides rich texture and color information but fails in darkness and gives no direct range. Lidar provides precise range and works in darkness but is expensive and blind to transparent surfaces. GPS gives global position but drifts indoors. Combining sensors — sensor fusion — is how real robots achieve robust perception.
Flashcards — click each card to reveal the answer
Noise, Resolution, and Sampling
Every physical sensor introduces imperfections. Understanding them quantitatively is essential — a control algorithm that ignores sensor noise will produce a jittery, unstable robot. Noise is random variation in the sensor's output that does not correspond to real changes in the measured quantity. Thermal noise in electronic circuits, photon shot noise in cameras, and vibration-induced jitter in accelerometers are all examples. Noise is often modeled as additive Gaussian: the true measurement is x, and the sensor reports x + n where n is drawn from a normal distribution with zero mean and standard deviation sigma. The parameter sigma is called the noise standard deviation or sensor noise floor. Resolution is the smallest change in the measured quantity that the sensor can detect. A 12-bit ADC (analog-to-digital converter) divides its input range into 2^12 = 4,096 discrete levels. If the input range is 0–5 V, the resolution is 5/4096 ≈ 1.22 mV — the sensor cannot distinguish inputs closer together than about 1.22 mV. Sampling rate is the frequency at which the sensor delivers measurements. The Nyquist-Shannon sampling theorem states that to faithfully capture a signal with frequency components up to f_max Hz, you must sample at least at 2 * f_max Hz. A robot wheel spinning at up to 10 revolutions per second generates an encoder signal with components up to 10 Hz; you need at least 20 samples per second to reconstruct its motion correctly. Undersampling — sampling too slowly — causes aliasing: high-frequency motion appears as spurious low-frequency signal.
Complete the key statements about sensor characteristics.
Filtering Raw Signals
Raw sensor data is almost never fed directly to a control algorithm. It is first processed by filters that reduce noise while preserving the signal of interest. The moving average filter computes the mean of the last N samples. It smooths out high-frequency noise but also blurs rapid legitimate changes — if a robot suddenly hits a wall, the moving average takes N steps to fully reflect the new sensor value. This trade-off between smoothness and responsiveness is fundamental to filter design. The low-pass filter (in continuous or discrete implementations) allows low-frequency signal components to pass while attenuating high-frequency components. A first-order low-pass filter is characterized by its cutoff frequency f_c: components below f_c pass with near-full amplitude, while components above f_c are progressively attenuated. A gyroscope's output contains genuine rotation signal at low frequencies and high-frequency vibration noise; a low-pass filter with f_c set below the vibration frequency recovers the useful signal. The Kalman filter is more sophisticated: it is a recursive algorithm that maintains a probabilistic estimate of the system state (position, velocity, etc.) and updates that estimate each time a new measurement arrives, optimally weighting each measurement by its uncertainty. A GPS update has high uncertainty (meters); a wheel encoder update has lower uncertainty over short intervals but drifts over time; a Kalman filter automatically blends both to maintain the best possible position estimate. Kalman filtering is the mathematical backbone of most production robot localization systems. Choosing the right filter requires understanding the signal frequency content, the noise characteristics, and the latency tolerance of the downstream algorithm.
A robot's IMU accelerometer reports a position drift of several centimeters per second even when the robot is stationary. What is the most likely cause of this behavior?
A wheel encoder measures shaft rotation at 50 Hz. The robot's wheels can spin at up to 30 revolutions per second. What problem will the engineer observe?
Sensor Selection for a Delivery Robot
- You are designing the sensor suite for a sidewalk delivery robot that must navigate city sidewalks, avoid pedestrians, and find building entrances autonomously.
- Step 1: List at least five sensors you would include. For each, specify: the sensor type, what it measures, its approximate operating range, and one key limitation.
- Step 2: Identify one proprioceptive sensor and one exteroceptive sensor from your list. Explain what information each provides and why both are necessary.
- Step 3: For each sensor, identify the most significant noise or error source it will experience in the sidewalk environment (vibration, rain, direct sunlight, GPS shadowing from buildings, etc.).
- Step 4: Propose a sensor fusion strategy: which two sensors would you combine first to address a specific limitation of each? Describe, at the conceptual level, how you would weight their outputs.
- Step 5: The robot's budget is cut and you must remove one sensor. Which do you remove, and what operational capability do you lose? Is the loss acceptable for the mission?