How Robots Sense

1. The Perceptual Gap

Robots have no direct awareness of true physical reality. Ground-truth coordinates are hidden state variables. Instead, physical transducers generate raw electrical voltages, photodiode currents, and digital register counts. The robot must continuously solve an inverse problem: inferring true physical states from noisy, indirect sensory evidence.

2. Proprioception and Encoders

Proprioceptive sensors measure internal bodily states. Optical wheel encoders feature a slotted disk rotating between an infrared emitter and detector. With radius $r = 0.10\text{ m}$ and $N = 1000\text{ ticks/rev}$, each pulse corresponds to a discrete increment of wheel rotation:

$\Delta x = 2\pi r \left(\frac{\Delta \text{ticks}}{N}\right) = 2\pi (0.10) \left(\frac{1}{1000}\right) \approx 0.628\text{ mm}$

A total count of $1592\text{ ticks}$ indicates exactly $1.00\text{ m}$ of wheel rotation.

3. Dead Reckoning Integration

To estimate continuous trajectory, the robot performs dead reckoning: summing incremental displacements over successive sampling intervals:

$x_k = x_{k-1} + v_k \Delta t$

If the wheel turns without slipping, integrating four successive $0.25\text{ m}$ displacements advances the internal position estimate smoothly from $x_0 = 0.00\text{ m}$ to $x_{\text{est}} = 1.00\text{ m}$.

4. The Wheel Slip Defect

Odometry assumes pure rolling contact without slipping. When the robot traverses slick wet tile with slip ratio $s = 0.15$ ($15\%$ slip), the wheels rotate through $1.00\text{ m}$ of circumference ($1592\text{ ticks}$), but the chassis translates only $d_{\text{actual}} = (1 - s) \times 1.00 = 0.85\text{ m}$. Dead reckoning accumulates an unrecoverable drift of $+0.15\text{ m}$.

5. Exteroception and Rangefinding

Exteroceptive sensors observe features in the external environment. A Time-of-Flight (ToF) laser rangefinder emits a short light pulse toward a known landmark wall located at $x_{\text{wall}} = 2.00\text{ m}$. By timing the photon return flight, the sensor measures a distance of $z = 1.15\text{ m}$. The robot calculates its absolute position directly:

$x_{\text{true}} = x_{\text{wall}} - z = 2.00 - 1.15 = 0.85\text{ m}$

Exteroception exposes and eliminates the $0.15\text{ m}$ odometry error.

6. Sensor Fusion

Neither sensor family alone is sufficient. Wheel encoders provide smooth, high-frequency updates ($50\text{--}100\text{ Hz}$) with negligible latency, but accumulate boundless drift. Range sensors anchor the robot to fixed physical landmarks, but update more slowly ($10\text{ Hz}$) and suffer occlusions. Sensor fusion algorithms (such as the Kalman filter) combine both streams to produce high-rate, drift-free state estimates.

7. Inertial Measurement Units

Inertial Measurement Units (IMUs) measure angular rates with gyroscopes and specific forces with accelerometers. However, MEMS accelerometers carry constant bias errors $b_a$. Double-integrating acceleration to obtain position causes the error to grow quadratically over time:

$\Delta x_{\text{drift}} = \frac{1}{2} b_a t^2$

Even a tiny bias of $b_a = 0.05\text{ m/s}^2$ causes $2.50\text{ m}$ of phantom drift in just $10\text{ seconds}$.

8. Stereo Vision and Triangulation

Stereo vision mimics human depth perception using two calibrated cameras separated by baseline distance $B = 0.12\text{ m}$. By identifying corresponding pixels between the left and right images, the horizontal disparity $d = x_L - x_R$ yields metric depth through similar triangles:

$Z = \frac{f \cdot B}{d}$

With focal length $f = 600\text{ px}$ and disparity $d = 60\text{ px}$, the target mug is located at depth $Z = (600 \times 0.12) / 60 = 1.20\text{ m}$.

9. Gaussian Measurement Uncertainty

Physical sensors do not output single true scalars; repeated observations scatter according to a probability distribution. A typical distance sensor measurement is modelled as true value plus zero-mean Gaussian noise: $z = x_{\text{true}} + v$, where $v \sim \mathcal{N}(0, \sigma^2)$. With standard deviation $\sigma = 0.05\text{ m}$, $68\%$ of raw readings fall within $1.15\text{ m}$ and $1.25\text{ m}$.

10. Where It Breaks

Sensors rely on physical interaction assumptions that the environment can violate. When a laser rangefinder beam strikes an angled glass pane or mirror, specular reflection bounces the light pulse away into the ceiling rather than returning an echo. The receiver detects zero return energy, prompting the robot to conclude the path is entirely clear and drive directly into the solid glass barrier.

The Quick Version

• Robots infer hidden physical reality from indirect digital sensor measurements.
• Optical encoders count discrete rotational pulses from slotted wheel disks.
• Dead reckoning integrates velocity increments over time to estimate trajectory.
• Wheel slip creates unrecoverable integration drift between encoder belief and reality.
• Exteroceptive LiDAR bounces energy off external landmarks to measure absolute clearance.
• Sensor fusion combines high-rate relative odometry with drift-free exteroceptive fixes.
• Accelerometer bias compounds quadratically into massive position errors under double integration.
• Stereo camera pairs compute metric 3D depth from horizontal pixel disparity.
• Raw sensor samples follow Gaussian probability distributions around true physical values.
• Specular glass barriers deflect laser pulses away, inducing blind physical collisions.