Sensor Noise

Gaussian noise, or white noise, is the most fundamental statistical model, characterized by a normal (bell-curve) distribution where each sample is independent and identically distributed. It is defined by its mean (often zero) and standard deviation (magnitude).

• Additive White Gaussian Noise (AWGN) is added directly to the signal.
• Models thermal noise in electronic circuits and fundamental quantum limits.
• In simulation, it's applied to sensor readings like camera pixel values, IMU measurements, and depth sensor returns to simulate ubiquitous electronic interference.

Salt-and-pepper noise is an impulse noise model where random pixels in an image (or samples in a signal) are set to either minimum (pepper, black) or maximum (salt, white) intensity values.

• Caused by dead pixels, bit errors in transmission, or sudden sensor malfunctions.
• In LiDAR or depth camera simulation, it manifests as sporadic, extreme outliers in the point cloud.
• Critical for testing a perception system's robustness to data corruption and the effectiveness of median filters.

Quantization noise arises from the fundamental process of converting a continuous analog signal into a discrete digital value with finite bit-depth. The error is the difference between the true analog value and its digital representation.

• Modeled as a uniform distribution over the range of one least significant bit (LSB).
• Directly impacts encoder resolution and the precision of analog-to-digital converters (ADCs).
• In simulation, it's crucial for accurately modeling the limited precision of real-world sensors like joint encoders and low-cost IMUs.

Bias noise is a systematic, non-zero mean error that is constant or slowly drifting over time. Unlike random noise, it represents a consistent offset in the sensor's output.

• Caused by sensor miscalibration, manufacturing imperfections, or environmental drift (e.g., temperature).
• Pervasive in IMU simulation, where accelerometers and gyroscopes have explicit bias parameters that must be modeled and later estimated via sensor calibration or a Kalman filter.
• A key challenge for long-term odometry and state estimation.

Temporal noise, such as 1/f noise (pink noise), has a power spectral density inversely proportional to frequency. This means lower-frequency components are stronger, causing the noise to appear as slow drifts or "flicker" in the signal.

• More realistic than white noise for many physical processes where samples are temporally correlated.
• Observed in inertial sensors, tactile sensors, and environmental sensors.
• Simulating it requires more complex time-series models (e.g., autoregressive processes) to test a control system's ability to reject low-frequency disturbances.

Shot noise originates from the discrete, quantized nature of energy or particles (e.g., photons, electrons). Its magnitude follows a Poisson distribution, where the variance equals the mean signal intensity.

• Dominant in imaging sensors (CMOS/CCD) and photon-counting LiDAR.
• In a camera simulation, it means pixel noise is signal-dependent: brighter regions have more absolute noise but a better signal-to-noise ratio.
• Accurately modeling shot noise is essential for simulating low-light conditions and evaluating high-dynamic-range (HDR) perception algorithms.

The most fundamental statistical noise model, Additive White Gaussian Noise (AWGN) adds random values drawn from a Gaussian (normal) distribution to a sensor's true signal. It models ubiquitous thermal noise in electronic circuits.

• Key Properties: Zero mean, constant power spectral density ('white'), and Gaussian amplitude distribution.
• Simulation Use: Applied to raw sensor readings (e.g., voltage from an encoder, pixel intensity) before any processing.
• Example: Adding AWGN to a simulated IMU's accelerometer reading to mimic the small, random vibrations present in any real electrical system.

Unlike random noise, bias is a constant offset error, and drift is a slow, time-varying change in that offset. These are systematic errors inherent to physical sensor manufacturing.

• Bias: A fixed value added to all measurements (e.g., a gyroscope that always reads 0.1 rad/s even when stationary).
• Drift: Bias that changes slowly over time, often modeled as a random walk or integrated Gaussian noise.
• Critical for Long-Term Tasks: Policies for navigation or manipulation must learn to be invariant to these slow-changing errors, which cannot be filtered out in a single timestep.

Quantization noise arises from the analog-to-digital conversion process, where a continuous signal is mapped to a finite set of discrete digital values. This creates a structured error bound by the sensor's resolution.

• Mechanism: The difference between the true analog value and the nearest representable digital value.
• Modeling: Often approximated as uniform noise over the interval ±(LSB/2), where LSB is the value of the least significant bit.
• Impact: Particularly important for simulating low-resolution sensors like inexpensive encoders or 8-bit depth cameras, where the step-like signal can significantly impact control stability.

Real sensor noise is often temporally correlated, meaning the error at one timestep depends on previous errors. Pink noise (1/f noise) and Brownian (red) noise model this 'colorful' reality.

• Pink Noise (1/f): Common in many electronic and environmental sensors. Power decreases with frequency.
• Red/Brownian Noise: Even more low-frequency dominated, modeling slow drifts and environmental trends.
• Simulation Method: Generated by filtering white noise through an appropriate low-pass filter. This creates more realistic, smoother error sequences that challenge naive filtering algorithms.

Sensors occasionally produce catastrophic errors or dropouts—single readings that are wildly incorrect. Modeling these outliers is crucial for robust perception.

• Salt-and-Pepper Noise: Random pixels in an image are set to either minimum (pepper) or maximum (salt) intensity.
• Broken Beam Model: For LiDAR/IR sensors, simulating the occasional missed return or phantom reflection.
• Purpose: Forces the downstream perception stack or policy to develop robustness, such as using median filters or RANSAC-based algorithms that can ignore spurious data points.

Each sensor modality has unique noise characteristics that must be modeled for high-fidelity simulation.

• Camera: Photon shot noise (Poisson-distributed), dark current noise, blooming, and lens-specific radial distortion.
• LiDAR: Beam divergence, mixed-pixel effects at edges, atmospheric attenuation, and multipath reflections.
• IMU: Scale factor nonlinearity, axis misalignment, g-dependent bias (for accelerometers), and vibration rectification.
• Tactile/Force: Hysteresis and nonlinear calibration curves. Simulating these profiles ensures a policy's robustness is tested against the specific failure modes of the hardware it will use.

Sensor fusion is the computational process of combining data from multiple, often noisy, sensors to produce a more accurate and reliable state estimate than any single sensor could provide. It is the primary method for mitigating the effects of individual sensor noise.

• Common Algorithms: Kalman Filters, Particle Filters, and Bayesian Networks.
• Purpose: Reduces uncertainty, fills in sensor blind spots, and increases system robustness.
• Example: Fusing a noisy GPS signal (low precision, high latency) with a noisy IMU (high frequency, drifting) to get a smooth, accurate position estimate.

A Kalman filter is an optimal recursive algorithm for estimating the state of a linear dynamic system from a series of noisy measurements. It is a foundational tool for sensor fusion and noise filtering.

• Mechanism: Iteratively predicts the system's next state using a model, then updates this prediction with a weighted average of the new, noisy sensor measurement.
• Key Output: Provides not just an estimate, but also a covariance matrix representing the uncertainty of that estimate.
• Use Case: Extensively used in navigation (GPS/INS), tracking, and robotics to filter out sensor noise like Gaussian noise from IMUs.

Sensor calibration is the process of determining the accurate intrinsic and extrinsic parameters of a sensor to correct for systematic errors, which are distinct from random noise.

• Intrinsic Calibration: Corrects for sensor-specific biases and scaling errors (e.g., camera lens distortion, IMU bias).
• Extrinsic Calibration: Determines the precise position and orientation of a sensor relative to a robot's body or other sensors.
• Relation to Noise: Calibration removes predictable, repeatable errors, leaving behind the unpredictable random noise that must be statistically modeled.

In simulation, ground truth refers to the perfectly accurate, noise-free data about the state of the simulated world that is known by the simulator itself.

• Purpose: Serves as the benchmark for evaluating the performance of perception algorithms and the effectiveness of noise models.
• Comparison: The difference between a sensor's noisy reading (e.g., a LiDAR point cloud) and the ground truth object pose quantifies the error introduced by the sensor model, including its noise.
• Critical Role: Essential for training and validating models in sim-to-real pipelines, as it provides the 'correct answer' that is unavailable in the real world.

Domain randomization is a sim-to-real technique where simulation parameters—including sensor noise characteristics—are deliberately varied across training episodes to force a policy to learn robust behaviors.

• Application to Noise: Instead of using a fixed noise model (e.g., Gaussian with σ=0.1), parameters like noise variance, bias, and dropout probability are sampled from a wide range for each training run.
• Goal: Prevents the policy from overfitting to the specific quirks of the simulation and encourages it to focus on invariant features, improving transfer to the noisy real world.
• Example: Randomizing the intensity of visual sensor noise, the drift of an IMU, and the error in joint position encoders during robotic arm training.

An actuator model is a mathematical representation of a physical motor's dynamics, which includes its own sources of noise and disturbance that affect control.

• Components: Models command-response latency, saturation limits, torque ripple, and electrical noise.
• Interaction with Sensors: Noisy actuators create vibrations and uncommanded motions that are then measured by proprioceptive sensors (e.g., joint encoders, IMUs), creating a coupled noise loop.
• Simulation Importance: A high-fidelity actuator model, complete with noise and non-linearities like friction, is necessary to train controllers that will work on physical hardware where perfect torque control is impossible.