Odometry

Wheel odometry estimates a robot's pose by integrating the rotational motion of its wheels, typically measured by rotary encoders. It is the most common method for ground-based robots.

• Mechanism: Encoders count wheel revolutions; the robot's kinematics model translates this into linear and angular displacement.
• Pros: High-frequency updates, low computational cost, and works indoors without external signals.
• Cons: Susceptible to wheel slip and uneven terrain, leading to unbounded drift over time. Accuracy degrades without frequent correction from other sensors.

Visual Odometry (VO) estimates egomotion by analyzing the apparent motion of visual features in a sequence of images from one or more cameras.

• Mechanism: Tracks feature points (e.g., using KLT tracker or ORB features) between consecutive frames and solves for camera motion that minimizes reprojection error.
• Types: Monocular VO (scale ambiguity) and Stereo VO (provides scale from baseline).
• Pros: Rich environmental data, no drift from wheel slip, usable in environments without predefined infrastructure.
• Cons: Computationally intensive, sensitive to lighting changes, fast motion, and low-texture environments (e.g., blank walls).

Visual-Inertial Odometry (VIO) is a sensor fusion technique that tightly couples visual data from a camera with inertial data from an Inertial Measurement Unit (IMU).

• Mechanism: The IMU provides high-frequency acceleration and angular velocity, bridging gaps in visual tracking and providing scale for monocular systems. Algorithms like MSCKF or optimization-based VINS-Fusion perform joint estimation.
• Pros: Highly robust to rapid motions and temporary visual occlusion; provides metric scale and gravity-aligned orientation.
• Cons: Requires careful sensor calibration (time synchronization, intrinsic/extrinsic parameters). Sensor biases must be estimated online.

LiDAR Odometry estimates motion by aligning consecutive 3D point cloud scans from a rotating laser rangefinder.

• Mechanism: Algorithms like Iterative Closest Point (ICP) or Normal Distributions Transform (NDT) find the rigid transformation that best aligns two point clouds.
• Pros: Highly accurate in structured environments (warehouses, streets); provides direct 3D structure; works in darkness.
• Cons: Expensive sensor; computationally heavy for dense point clouds; performance can degrade in featureless environments (e.g., long corridors). Often fused with IMU data for smoother estimates.

An Inertial Navigation System (INS) estimates pose by double-integrating linear acceleration and integrating angular rate from an IMU—a process known as dead reckoning.

• Mechanism: Accelerometer data is transformed into the navigation frame, corrected for gravity, and integrated twice to get position. Gyroscope data is integrated to get orientation.
• Pros: Self-contained, immune to external signal jamming or denial, provides very high-frequency pose estimates.
• Cons: Unbounded exponential drift due to sensor noise and bias integration. Standalone INS is unusable for precise navigation over more than a few seconds without external aiding (e.g., from GPS or VO).

No single odometry method is perfect for all conditions. Sensor fusion combines multiple, complementary sources to create a robust, high-integrity state estimate.

• Common Fusion Architectures:
  • Filter-based: Extended Kalman Filter (EKF) or Particle Filter fuse wheel, IMU, and occasional absolute measurements (e.g., from beacons or loop closure).
  • Optimization-based: Factor graph frameworks (e.g., GTSAM) fuse heterogeneous measurements as probabilistic constraints for batch or incremental smoothing.
• Key Benefit: Mitigates the weaknesses of individual sensors (e.g., IMU drift corrected by visual data; visual tracking gaps filled by IMU). This is critical for heterogeneous fleets operating in dynamic, unpredictable environments.

Visual Odometry estimates a robot's egomotion by analyzing the sequence of images from onboard cameras. Instead of wheel encoders, it tracks visual features across frames to compute pose changes.

• Primary Sensor: Monocular or stereo cameras.
• Key Technique: Feature matching and optical flow.
• Challenge: Sensitive to lighting, texture, and motion blur.
• Use Case: Camera-equipped robots in GPS-denied, structured indoor environments.

VIO fuses camera data with inertial measurements from an IMU to create a robust, high-frequency pose estimate. The IMU provides high-rate motion data between camera frames, smoothing estimates and handling aggressive motions.

• Sensor Fusion: Combines visual feature tracking with accelerometer and gyroscope data.
• Benefit: More robust than VO alone, especially during rapid turns or low-texture scenes.
• Common Framework: Popular implementations include VINS-Mono and OKVIS.

Dead reckoning is the foundational principle behind odometry: estimating current position by advancing a previous position based on measured velocity and elapsed time.

• Core Mechanism: Integration of velocity or wheel encoder ticks over time.
• Inherent Limitation: Unbounded accumulation of error (drift) due to sensor noise and wheel slip.
• Contrast with Odometry: Often used interchangeably, but odometry typically refers to the specific implementation using proprioceptive sensors (encoders, IMU), while dead reckoning is the broader mathematical concept.

An IMU is a key sensor for odometry, providing raw inertial data. It typically contains a 3-axis accelerometer and a 3-axis gyroscope.

• Measurements: Specific force (acceleration minus gravity) and angular rate.
• Role in Odometry: Used in wheel-IMU odometry to correct for wheel slip and provide orientation estimates.
• Characteristic: High-frequency data but suffers from significant bias and noise drift, requiring careful calibration and filtering.

Drift is the fundamental error in odometry: the unbounded accumulation of small measurement errors over time, causing the estimated position to deviate from the true position.

• Primary Cause: Integration of noisy sensor data (encoder ticks, IMU readings).
• Secondary Causes: Wheel slippage, uneven floors, and mechanical imperfections.
• Mitigation: Odometry is always fused with other sensing (e.g., LiDAR, cameras) via sensor fusion algorithms like the Kalman Filter or via loop closure in a SLAM system to correct drift periodically.

A motion model (or process model) is the mathematical function that predicts how a robot's state (pose) changes given its control inputs. It is the core predictive component in state estimation filters.

• For Wheeled Robots: Often a kinematic model (e.g., differential drive, Ackermann steering).
• Input: Wheel encoder velocities or motor commands.
• Output: Predicted change in position (dx, dy) and orientation (dθ).
• Usage: Provides the 'prediction' step in Bayesian filters like the Kalman Filter or Particle Filter, which then correct the prediction with sensor observations.