Mobile Robot Kinematics and Localization

Introduction

Mobile robotics concerns machines that move through an environment rather than remaining fixed to a workspace. Their motion is governed not only by actuator commands but also by wheel geometry, contact conditions, terrain, vehicle orientation, sensing uncertainty, and environmental constraints. A robot may know the velocity commanded to its motors while still being uncertain about its actual position. This distinction motivates the study of kinematics, odometry, localization, mapping, and simultaneous localization and mapping (SLAM).

A useful mobile-robot model has three layers. First, the configuration describes pose and, where necessary, joint or body orientation. Second, the kinematic model maps actuator motion to vehicle velocity. Third, the estimation system combines motion predictions with sensor measurements to infer pose and environmental structure. This chapter develops these layers for wheeled, legged, aerial, and underwater platforms.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Model differential-drive, Ackermann-steered, omnidirectional, legged, aerial, and underwater robots.
2. Distinguish holonomic and nonholonomic constraints.
3. Derive and use basic forward and inverse kinematic relationships.
4. Explain encoder odometry, its error sources, and uncertainty growth.
5. Compare landmark-based localization, filtering, mapping, and SLAM.
6. Select an appropriate localization architecture for an engineering application.
7. Evaluate practical, safety, ethical, and operational limitations of mobile-robot navigation.

1. Pose, Frames, and Configuration

In a planar environment, a rigid robot pose is commonly represented by

where x and y are Cartesian coordinates in a fixed world frame, and θ is the robot heading measured counterclockwise from the world x-axis. The robot body frame is usually attached to its center, with xᵦ pointing forward and yᵦ pointing laterally.

A body-frame velocity is written as ξ = [v_x, v_y, ω]^T, where vₓ and vᵧ are translational velocities and ω is angular velocity. For a planar robot,

The matrix rotates body-frame translational velocity into world coordinates. It assumes a rigid body, a planar workspace, and a consistent definition of θ. The inverse transformation is obtained by transposing the rotation matrix.

For a finite sampling interval Δt, a simple discrete approximation is

This Euler approximation is suitable when velocities are nearly constant and Δt is sufficiently small. For significant angular motion, an exact constant-twist integration or a higher-order method is preferable.

A configuration space includes all independent variables required to describe the robot. A wheeled vehicle may have three planar pose variables but fewer instantaneous velocity freedoms because wheel contacts impose constraints. This difference between configuration dimension and instantaneous mobility is fundamental.

2. Holonomic and Nonholonomic Constraints

A constraint is holonomic if it can be expressed solely as a relation among configuration variables, such as f(q) = 0. A constraint is nonholonomic when it restricts velocities and cannot generally be integrated into a constraint involving configuration alone.

A holonomic planar robot can independently command vₓ, vᵧ, and ω. Such a robot is often called omnidirectional. A conventional car cannot move directly sideways: its instantaneous velocity is aligned with its longitudinal axis. Its no-slip constraint is

This equation states that lateral velocity in the body frame is zero. It is a velocity constraint and therefore nonholonomic. A nonholonomic robot may nevertheless reach many positions through sequences of admissible motions; it simply cannot realize every instantaneous velocity.

The distinction has practical consequences. A holonomic robot can track a desired planar velocity directly, while a nonholonomic robot must generate a feasible trajectory, often involving arcs, turns, and forward motion. Motion planners must therefore respect the robot's constraint model rather than treating every platform as a point mass.

3. Differential-Drive Robots

A differential-drive robot has two independently driven wheels separated by track width b. Let the wheel radius be r, and let right and left wheel angular speeds be ωᵣ and ωₗ. Under pure rolling and no lateral slip, the wheel linear speeds are rωᵣ and rωₗ. The body forward velocity and angular velocity are

The forward kinematic model is therefore

or, equivalently,

The inverse model, useful for control, is

These equations assume equal effective wheel radii, a rigid axle, pure rolling, negligible compliance, and accurate wheel separation. If both wheels have equal speed, the robot moves straight. If their speeds are equal and opposite, it rotates approximately about its midpoint. If one wheel is stationary, the robot follows a circular path about that wheel.

For constant v and ω ≠ 0, the instantaneous center of curvature lies at signed distance R = v/ω from the body center. Over a short interval, the robot follows an arc rather than a straight line. When |ω| is very small, the straight-line limit should be used numerically to avoid division by a small number.

Worked Example: Differential Drive

A robot has r = 0.10 m, b = 0.50 m, ωᵣ = 6 rad/s, and ωₗ = 4 rad/s. Then

v = 0.10(6 + 4)/2 = 0.50 m/s

and

ω = 0.10(6 − 4)/0.50 = 0.40 rad/s.

Thus the robot moves forward at 0.50 m/s while turning counterclockwise at 0.40 rad/s. Its instantaneous turning radius is R = 0.50/0.40 = 1.25 m. The calculation assumes no slip and uses radian measure, for which radians are dimensionless in the velocity equations.

4. Ackermann-Steered Vehicles

An Ackermann vehicle approximates a car or mobile robot with front steering and rear-wheel drive or passive rear wheels. Its idealized bicycle model has wheelbase L, forward speed v, and steering angle δ. The kinematic equations are

The curvature is κ = tanδ/L, and the turning radius is R = L/tanδ when δ is nonzero. The model assumes low-speed motion, negligible tire slip, a single equivalent front wheel, and steering geometry represented by δ.

Real Ackermann steering uses different inner and outer wheel angles so that the wheel axes intersect at a common instantaneous center. If the distance from the rear axle to the instantaneous center is R, and the front-track width is T, ideal steering angles satisfy

The signs and “inside/outside” assignment depend on the turning direction. At higher speeds, tire slip, load transfer, suspension compliance, and steering dynamics make the kinematic model insufficient; a dynamic bicycle model is then required.

5. Omnidirectional Platforms

Omnidirectional robots can generate planar motion in more than one direction without first reorienting their body. Common mechanisms include Swedish or mecanum wheels and three-wheel or four-wheel platforms with angled rollers.

For a general wheeled platform, each wheel imposes a rolling-direction relationship. A compact velocity model is

where ξ = [vₓ, vᵧ, ω]ᵀ is the body twist, u is the vector of wheel speeds, and J is determined by wheel locations, wheel radii, and roller orientations. If J has rank three, the platform can command all three planar velocity components locally and is holonomic in the plane.

For a three-wheel omni platform with wheel radius r, wheel position angle αᵢ, and wheel rolling direction arranged tangentially, one commonly used convention gives

where l is the distance from the body center to wheel i. The exact signs depend on the chosen wheel-axis convention; the geometry must be defined before implementation. The inverse relationship is obtained by solving the stacked wheel equations, using a matrix inverse when square and nonsingular or a pseudoinverse when redundant.

Mecanum wheels provide apparent omnidirectional motion through angled rollers. Their performance is sensitive to roller geometry, floor flatness, wheel-load distribution, and slip. A mecanum robot may be mathematically holonomic while experiencing substantial practical odometry error on dusty, uneven, or compliant surfaces.

6. Legged, Aerial, and Underwater Platforms

Legged Robots

Legged robots do not have a continuously rolling contact constraint. Their motion alternates among support phases, such as standing, single support, and double support. During a support phase, the feet in contact with the ground are modeled as constraints. A stationary foot may impose zero velocity at its contact point:

The body and joint velocities are related through a Jacobian,

where q contains body and joint coordinates. A legged robot is dynamically and kinematically hybrid: the equations change when contacts are created or broken. Foot slippage, terrain uncertainty, compliance, and impact make state estimation more difficult than encoder integration alone. In practice, inertial measurement units (IMUs), joint encoders, foot-contact detection, cameras, and force sensors are fused to estimate body motion.

Aerial Robots

Aerial vehicles generally have six degrees of freedom: three translations and three rotations. Their pose belongs to the rigid-body configuration space SE(3). A body-frame translational velocity v and angular velocity ω are transformed to the world frame using a rotation matrix R:

where p is world position. Attitude may be represented by Euler angles, rotation matrices, or unit quaternions. Euler angles are intuitive but can suffer singularities; quaternions avoid this representation singularity but require normalization. A quadrotor is often treated as an underactuated system: four rotor thrusts control six pose variables, with attitude used to orient the total thrust vector.

GPS, IMUs, cameras, lidar, and barometers are commonly fused. Indoor flight requires alternatives to GPS, such as visual-inertial odometry, lidar-inertial estimation, motion-capture systems, or ultra-wideband ranging.

Underwater Vehicles

Underwater robots operate in six degrees of freedom and are affected by buoyancy, hydrodynamic drag, added mass, current, and pressure. Their kinematics can still be expressed as

but the dynamics are strongly environment-dependent. GPS does not work underwater, so localization may use inertial navigation, Doppler velocity logs, acoustic transponders, sonar, magnetic sensors, terrain matching, or surfacing fixes. Acoustic measurements have limited bandwidth and may be affected by multipath propagation. This makes underwater SLAM and long-duration navigation particularly challenging.

7. Odometry and Its Uncertainty

Odometry estimates pose by integrating measured motion, typically from wheel encoders, joint encoders, IMUs, or visual feature tracking. For a differential-drive robot, encoder increments Δsᵣ and Δsₗ produce

A midpoint update is

Here Δsᵣ and Δsₗ are signed distances traveled by the right and left wheels during one sampling interval. The update assumes constant wheel increments over the interval and no slip. It is generally more accurate than evaluating translation only at θₖ when the robot turns during the interval.

Odometry is locally smooth and available at high rate, but it is not globally reliable. Systematic errors include unequal wheel radii, incorrect wheelbase calibration, encoder scale errors, and encoder index offsets. Nonsystematic errors include wheel slip, floor irregularity, collisions, and changing payload. Orientation error is especially damaging because it rotates subsequent translational estimates.

A state estimator represents uncertainty with a covariance matrix P. In a simplified linear prediction model,

where Fₖ is the state-transition Jacobian and Qₖ is process-noise covariance. The equation assumes a locally linearized model and zero-mean process noise. It does not imply that errors are truly Gaussian in every environment; covariance is an approximation whose quality depends on the model and noise assumptions.

8. Localization

Localization estimates a robot's pose relative to a known or partially known reference. Major categories are:

• Dead reckoning: integrates motion without external correction.
• Relative localization: estimates motion relative to a previous pose or local map.
• Absolute localization: uses landmarks, GPS, surveyed beacons, or map references.
• Global localization: determines pose when the initial pose may be unknown.
• Kidnapped-robot recovery: re-localizes after an unexpected displacement.

A range-bearing landmark observation can be written as

where mᵢ is landmark position, h is the nonlinear measurement function, and vᵢ is sensor noise. For a planar landmark at (mₓ, mᵧ), the ideal range and bearing are

The bearing must be wrapped to a chosen interval such as (−π, π]. A known map and reliable data association permit correction of odometric drift.

The extended Kalman filter (EKF) predicts a state and covariance, then uses a linearized measurement model H:

Here x̂⁻ and P⁻ are the predicted state and covariance, x̂⁺ and P⁺ are corrected values, R is measurement-noise covariance, and K is the Kalman gain. The EKF is appropriate when nonlinearities are moderate, uncertainty is reasonably represented by a single local Gaussian, and Jacobians are available. Particle filters are often preferable for multimodal or globally ambiguous localization.

9. Mapping and SLAM

A map may consist of point landmarks, line segments, occupancy-grid cells, a topological graph, or a three-dimensional point cloud. An occupancy grid divides space into cells with estimated occupancy probabilities. Log-odds are often used because repeated evidence can be accumulated additively:

where lₜ(m) is the log-odds for cell m at time t, zₜ is the sensor observation, xₜ is the pose, and l₀ is the prior log-odds. The inverse sensor model determines which cells are likely occupied or free. This formulation assumes a chosen sensor model and conditional independence approximation; real sensor errors are often correlated.

SLAM estimates the robot trajectory and map simultaneously. A conceptual probabilistic objective is

where x₀:T is the sequence of poses, m is the map, z are observations, and u are controls or odometric inputs. SLAM is difficult because pose uncertainty changes the interpretation of every measurement, while the map is needed to correct pose uncertainty.

Common approaches include:

• EKF-SLAM: maintains a joint Gaussian over robot pose and landmark locations; computational cost and linearization errors increase with map size.
• Particle-filter SLAM: represents multiple possible trajectories and is effective for some nonlinear, ambiguous problems, but may require many particles.
• Graph-based SLAM: represents poses as nodes and relative constraints as edges, then optimizes the graph. It is effective for large-scale mapping and loop closure.
• Visual-inertial and lidar-inertial SLAM: combine complementary sensing modalities to improve scale, robustness, and motion estimation.

A loop closure occurs when the robot recognizes a previously visited place. It adds a constraint connecting distant parts of the trajectory and can greatly reduce accumulated drift. False loop closures are dangerous because an incorrect global constraint can distort the entire map; robust data association and verification are therefore essential.

10. Worked Localization Example

Suppose a differential-drive robot starts at q₀ = (0, 0, 0), has wheelbase b = 0.40 m, and records wheel travel Δsᵣ = 0.12 m and Δsₗ = 0.08 m during one interval. Then

Δs = (0.12 + 0.08)/2 = 0.10 m

and

Δθ = (0.12 − 0.08)/0.40 = 0.10 rad.

Using the midpoint update,

x₁ = 0.10 cos(0.05) ≈ 0.0999 m,

y₁ = 0.10 sin(0.05) ≈ 0.0050 m,

θ₁ = 0.10 rad, approximately 5.73°.

The robot has moved almost 0.10 m forward while turning slightly left. The nonzero y-coordinate results from integrating motion along the changing heading rather than assuming a perfectly straight path.

11. Applied Case: Warehouse Mobile Robot Fleet

Background

Consider an illustrative case involving autonomous differential-drive robots transporting containers between storage racks and packing stations. The warehouse contains repetitive shelving, smooth concrete, pedestrians, reflective surfaces, and occasional wheel contamination.

Problem

Wheel odometry is available at 100 Hz, but its error grows during long journeys. Lidar measurements provide geometric information but may be ambiguous in repeated aisles. The system must maintain localization, stop safely near people, and recover after temporary sensor obstruction.

Analysis

The differential-drive model provides high-rate control and prediction. However, wheel slip during acceleration and variations in effective wheel radius cause systematic and random error. A map-based lidar system can correct drift, but repeated rack geometry creates data-association ambiguity. A single estimate without uncertainty is inadequate because navigation decisions depend on confidence.

Approach and Solution

The robot uses calibrated encoder odometry as the motion model and an IMU to improve short-term heading prediction. A lidar-based scan matcher supplies relative pose constraints. A pose-graph SLAM or localization backend incorporates verified loop closures and known warehouse landmarks. The local planner uses the nonholonomic model, while a safety layer independently monitors lidar and emergency-stop sensors.

Localization quality is assessed using innovation consistency, trajectory error against surveyed reference points, loop-closure residuals, and failure-recovery time. Maps are version-controlled so that changes in racks or temporary obstacles do not silently corrupt the permanent reference map. A fleet manager rejects navigation goals when localization covariance exceeds a safety threshold.

Outcome

The combined system is more reliable than odometry alone because drift is periodically corrected. It is more robust than scan matching alone because the motion prior limits implausible correspondences. Safety remains independent of the mapping algorithm: a robot must stop when obstacle sensing is uncertain, even if its estimated pose appears accurate.

Learning Points

1. No single sensor is reliable in every warehouse condition.
2. Kinematic feasibility must be included in planning and control.
3. Uncertainty should affect operational decisions, not merely be displayed.
4. Loop closure requires verification because false matches can be catastrophic.
5. Safety monitoring should remain effective during localization or software failure.

12. Design Practice, Limitations, and Professional Issues

A sound mobile-robot design begins with a frame convention, units, sign definitions, and calibration procedure. Engineers should validate kinematic models with straight-line, in-place rotation, and known-radius tests. Encoder measurements should be time-stamped and synchronized with IMU and range data. Angles must be normalized consistently, and coordinate transformations must be tested using known poses.

Model choice should match the platform and operating regime. A kinematic car model is useful at low speed but cannot predict tire-force effects at high speed. Wheel odometry is excellent for short-term relative motion but cannot guarantee global position. GPS provides an absolute reference outdoors but may be unavailable or unsafe near structures. Lidar performs well in geometric environments but can struggle with feature-poor or dynamic scenes. Cameras provide rich information but are affected by lighting, blur, texture, and privacy concerns.

Dynamic obstacles create a distinction between a static map and the current world. A map should not be treated as an unquestionable representation of reality. Perception systems may misclassify people, animals, or unusual objects. Safety standards, risk assessment, emergency stopping, human override, cybersecurity, and operational logging are engineering requirements, not optional additions.

The use of AI-based perception and learned localization introduces additional concerns: training-data bias, uncertain behavior outside the training distribution, explainability, adversarial interference, and protection of recorded images or facility maps. Human monitoring and conservative fallback behavior are especially important in public, industrial, medical, and transportation environments.

Contemporary systems increasingly combine lidar, cameras, IMUs, wheel or leg odometry, GNSS, radar, and learned feature representations. Improvements in factor-graph optimization, event cameras, solid-state lidar, neural implicit maps, and cloud-assisted fleet localization are significant, but they do not eliminate the need for calibrated models, observability analysis, uncertainty estimation, and safety validation.

Chapter Summary

Mobile-robot kinematics relates actuator motion and contact geometry to platform velocity. Differential-drive robots use wheel-speed sums for translation and differences for rotation. Ackermann vehicles obey a steering-dependent curvature relationship and are intrinsically nonholonomic. Omnidirectional mechanisms can command planar translation and rotation, although real slip may limit their ideal behavior. Legged robots have changing contact constraints, while aerial and underwater robots require six-degree-of-freedom pose models and specialized sensing.

Holonomic constraints depend only on configuration; nonholonomic constraints restrict admissible velocities. Odometry integrates measured motion but accumulates uncertainty, especially through wheel slip and heading error. Localization corrects motion estimates using landmarks, maps, or absolute references. Mapping represents environmental structure, and SLAM estimates both map and trajectory when neither is initially known. EKF, particle-filter, and graph-based methods each make different assumptions and have different scalability and robustness properties.

Professional systems combine complementary sensors, quantify uncertainty, verify data association, respect vehicle constraints, and retain independent safety mechanisms. The central engineering principle is that a robot should not merely calculate a pose; it should know how much that pose can be trusted and act accordingly.

Key Terms

Pose: Position and orientation of a robot in a reference frame.

Configuration: A complete set of variables describing the robot's state of arrangement.

Body frame: Coordinate frame attached to the robot.

World frame: Fixed external coordinate frame used for global representation.

Holonomic: Having independently commandable velocity directions subject to integrable constraints.

Nonholonomic: Subject to velocity constraints that cannot generally be integrated into configuration constraints.

Differential drive: Mobile base driven by independently actuated wheels on opposite sides.

Ackermann steering: Steering geometry that approximates automobile motion with a common instantaneous center.

Omnidirectional robot: Robot capable of planar motion in arbitrary instantaneous directions.

Odometry: Pose estimation obtained by integrating measured motion.

Wheel slip: Difference between assumed rolling motion and actual contact motion.

Localization: Estimation of robot pose relative to a reference or map.

Landmark: Distinct environmental feature used for estimation.

Occupancy grid: Discretized map whose cells carry occupancy estimates.

Data association: Matching sensor observations to landmarks or map features.

Loop closure: Recognition of a previously visited place in SLAM.

SLAM: Simultaneous localization and mapping.

Covariance: Matrix describing estimated uncertainty and correlation.

Pose graph: Graph of pose nodes connected by relative-measurement constraints.

Observability: Ability to infer state variables from available inputs and measurements.

Review Questions

Short Answer Questions

1. What is the difference between a body frame and a world frame?
2. Why is the no-sideways-motion condition of a car nonholonomic?
3. State the forward kinematic equations for a differential-drive robot.
4. Name two systematic and two nonsystematic sources of odometry error.
5. What is the purpose of loop closure in SLAM?

Long Answer Questions

1. Derive the inverse kinematic equations for a differential-drive robot and explain the assumptions behind them.
2. Compare differential-drive, Ackermann, and omnidirectional platforms in terms of motion constraints and applications.
3. Explain how an EKF combines a motion prediction with a landmark measurement.
4. Discuss why legged, aerial, and underwater robots require sensing strategies different from ordinary wheeled robots.
5. Explain the architecture and engineering trade-offs of a warehouse robot localization system.

Analytical / Application Questions

1. A robot has wheel radius 0.08 m, wheel separation 0.36 m, and wheel angular speeds 5 rad/s and 3 rad/s. Calculate its forward and angular velocities, state the turning direction, and determine its instantaneous turning radius.
2. Design a sensor-fusion strategy for an indoor aerial robot operating without GPS. Identify the motion model, sensors, major failure modes, and an appropriate estimator.
3. A lidar-based SLAM system produces a large map distortion after a false loop closure. Explain how the failure could be detected, what robust methods could reduce its likelihood, and how the system should recover safely.

Multiple Choice Questions with Answers

Question 1

Which equation describes the no-lateral-slip constraint of a planar car?

A. −sinθ ẋ + cosθ ẏ = 0

B. ẋ² + ẏ² = 1

C. θ̇ = 0

D. ẋ = ẏ

Correct Answer: A. −sinθ ẋ + cosθ ẏ = 0

Question 2

A differential-drive robot has wheel radius 0.1 m, wheel separation 0.5 m, and wheel speeds 6 rad/s and 4 rad/s. What are its forward and angular velocities?

A. 0.2 m/s and 0.4 rad/s

B. 0.5 m/s and 0.4 rad/s

C. 1.0 m/s and 2.0 rad/s

D. 0.4 m/s and 0.5 rad/s

Correct Answer: B. 0.5 m/s and 0.4 rad/s

Question 3

Which platform is generally holonomic in the plane when its wheel Jacobian has rank three?

A. Conventional Ackermann vehicle

B. Tracked vehicle with fixed tracks

C. Omnidirectional wheeled platform

D. Single-steered bicycle model

Correct Answer: C. Omnidirectional wheeled platform

Question 4

What is the primary weakness of dead-reckoning odometry?

A. It cannot provide high-rate estimates

B. Its uncertainty generally accumulates over time

C. It requires a pre-existing map

D. It cannot estimate orientation

Correct Answer: B. Its uncertainty generally accumulates over time

Question 5

In the differential-drive midpoint update, why is θk + Δθ/2 used?

A. It estimates the average heading during the interval

B. It removes wheel-radius errors

C. It forces the robot to move straight

D. It measures the wheelbase

Correct Answer: A. It estimates the average heading during the interval

Question 6

What does a loop closure provide in graph-based SLAM?

A. A new actuator command

B. A constraint connecting the current pose to a previously visited place

C. A guarantee that all landmarks are static

D. A replacement for sensor calibration

Correct Answer: B. A constraint connecting the current pose to a previously visited place

Question 7

Which estimator is particularly suited to representing multiple possible global poses?

A. Single linear least-squares estimate only

B. Particle filter

C. Fixed-gain tachometer

D. Pure wheel integration

Correct Answer: B. Particle filter

Question 8

For an Ackermann vehicle, what is the idealized curvature?

A. κ = L/tanδ

B. κ = tanδ/L

C. κ = δL

D. κ = 1/(L tanδ)

Correct Answer: B. κ = tanδ/L

Question 9

Why is GPS generally unavailable as a direct positioning sensor underwater?

A. Water amplifies GPS signals

B. GPS signals do not propagate usefully through water

C. Underwater robots have no orientation

D. GPS only measures angular velocity

Correct Answer: B. GPS signals do not propagate usefully through water

Question 10

Which condition most directly threatens the validity of a pure-rolling wheel model?

A. A calibrated encoder

B. A flat rigid floor

C. Wheel slip on a contaminated surface

D. A fixed coordinate convention

Correct Answer: C. Wheel slip on a contaminated surface