Kinematic Jacobian in Differential-Drive Robots

• Kinematic Jacobian is a mathematical mapping that converts differential-drive wheel velocities into linear and angular motions while considering nonholonomic constraints.
• It is derived from key geometric parameters such as wheel radius and axle length, and is fundamental to accurate state estimation and control.
• The Jacobian underpins advanced control strategies like MPC and spatiotemporal robust control, ensuring precise trajectory tracking and adaptability in dynamic environments.

A differential-drive robot is a mobile robotic system equipped with two independently actuated wheels mounted on a common axis, sometimes complemented by passive casters for stability. Differential drive architecture is foundational in robotics due to its mechanical simplicity, controllability, and effectiveness in planar navigation and manipulation tasks. The platform's characteristic property is that its motion—linear and angular velocity in the plane—can be fully and instantaneously modulated by adjusting the relative velocities of the left and right wheels, under the idealized assumptions of perfect wheel–ground contact and no lateral slip. More sophisticated variants, such as systems with omnidirectional wheels and variable base width (e.g., Omni Differential Drive), extend these principles while maintaining differential modulation as a core concept. The differential-drive model has been mathematically formalized to accommodate uncertainties, actuator limits, and time-varying environments, forming the basis for modern control, estimation, and multi-agent planning frameworks.

1. Mathematical Modeling and Kinematics

The canonical state representation for a differential-drive robot is $\xi = [x_1, x_2, \theta]^T$, where $x_1$ and $x_2$ denote planar position and $\theta$ is orientation. The standard control input is $u = [v, \omega]^T$, with $v$ as the forward velocity and $\omega$ as the angular velocity about the z-axis. The nonholonomic kinematic equations are: $$\dot{\xi} = \begin{bmatrix} v \cos \theta \ v \sin \theta \ \omega \end{bmatrix}$$ with additional dynamics or uncertainty terms as needed, e.g.,

$\dot{\xi} = f(\xi, u) + d(\xi, t)$

where $d(\xi, t)$ represents model uncertainties and external disturbances, typically assumed locally Lipschitz in $x_1$0 and uniformly bounded by $x_1$1 (Das et al., 5 Dec 2025).

Wheel velocities, $x_1$2 and $x_1$3, are mapped to body velocities using robot geometry: $x_1$4 where $x_1$5 is wheel radius and $x_1$6 is axle length (Ali et al., 2024). The discrete-time versions and shifted reference points (e.g., for external sensor fusion) are also widely used (Razak et al., 2019), and extensions incorporate nonzero lateral slip for tracked vehicles or skid-steer platforms via instantaneous center-of-rotation (ICR) formulations (Zhang et al., 2024).

2. Control Methodologies: Robustness, Optimality, and Adaptation

Differential-drive robots are controlled using a variety of advanced methods, each targeting specific performance goals such as robustness, time-optimality, safety, or adaptability.

• Spatiotemporal Tube (STT) Robust Control: An STT corresponds to a dynamically varying corridor, $x_1$7, with continuously differentiable center $x_1$8 and radius $x_1$9. For specification satisfaction under model uncertainty, an offline Scenario Optimization Program (SOP) samples time, enforces collision and reachability at those samples, and minimizes a slack $x_2$0 such that $x_2$1 guarantees the Temporal Reach-Avoid-Stay (T-RAS) property. A closed-form, approximation-free online control law then ensures state containment within $x_2$2 despite lumped uncertainty (Das et al., 5 Dec 2025).
• Model Predictive Control (MPC) with Barrier Functions: Dynamic feedback linearization (DFL) transforms the nonlinear system to a chain-of-integrators, enabling Linear MPC (LMPC) with quadratic costs and affine constraints. Safety is enforced with Control Barrier Functions (CBFs), $x_2$3, such that obstacle avoidance is guaranteed by maintaining $x_2$4 at each MPC step (Ali et al., 2024).
• Computed Torque and Gray-Box RL Hybridization: Computed Torque Control (CTC) leverages full dynamic models (including Coriolis, friction, and wheel inertia effects) for high-fidelity trajectory tracking. To address modeling inaccuracies, gray-box approaches integrate CTC structure with Data-Efficient Deep Reinforcement Learning, learning only uncertain parameters within physically plausible ranges and enforcing critically-damped closed-loop pole placement (Pishkhani, 30 Aug 2025).
• Adaptive and Distributed Coverage: For multi-robot scenarios, adaptive laws using online parameter estimation (e.g., adaptive L2-density matching) and consensus schemes improve convergence to optimal coverage configurations. The distributed control laws invert the differential-drive kinematic matrix and couple feedback to local Voronoi centroid estimates and learned density parameters (Razak et al., 2019).
• Lyapunov-Based and Noise-Robust Control: Point-to-point and trajectory tracking are achieved via discrete-time Lyapunov function construction, with tailored global-local splitting of operating regions. Gains are explicitly chosen to guarantee convergence and boundedness under system and measurement noise (Nguyen et al., 2020).

3. State Estimation and Sensor Fusion

Differential-drive robots require robust state estimation, particularly in the presence of partial observability, sensor noise, and drift. Multi-sensor fusion via the Extended Kalman Filter (EKF) integrates wheel encoders, IMU, visual and LiDAR odometry. State vectors typically include position, orientation, and velocities. Process and observation models reflect the robot’s kinematics and the characteristics of each sensor. In the case of failure or noise bursts (e.g., visual odometry dropouts), the EKF retains consistent state estimates by adjusting process and measurement noise covariances; partial updates allow asynchronous sensor streams (Alwala et al., 16 Mar 2026).

4. Trajectory Optimization and Planning

Trajectory generation for differential-drive robots must account for nonholonomic constraints, dynamic limits, and the specifics of the actuation mechanism.

• Universal Trajectory Optimization: This framework parametrizes trajectories via polynomials over motion states (e.g., angle, arc length) rather than Euclidean position, inherently matching the robot’s nonholonomic constraints. The optimization objective penalizes higher derivatives (e.g., jerk), maintains feasibility through a suite of safety and kinematic constraints, and leverages efficient augmented Lagrangian Newton-type solvers (Zhang et al., 2024).
• Time-Optimal Path Parameterization (TOPP-DWR): Given a geometric path, the speed profile along it is optimized to minimize traversal time while respecting bounds on linear/angular velocity, acceleration, and wheel rates. These constraints are reformulated as linear inequalities in the variables $x_2$5, and the resulting second-order cone programming (SOCP) problem is solved efficiently (≤1 ms in practice) (Li et al., 17 Nov 2025).
• Multi-Agent Coordination: For large-scale fleets, frameworks like MASS decouple multi-agent path planning (MAPF) from single-agent kinodynamic feasibility. The planning architecture leverages a three-level cascade: MAPF for high-level spatiotemporal coordination; Stationary State Search for per-agent kinodynamic trajectory synthesis; and LP or Bézier curve solvers for speed profile generation (Yan et al., 2024).

5. Singular Scenarios: Games, Surveillance, and Human Assistance

Differential-drive robots have been systematically studied in pursuit-evasion games and in assistive scenarios where nonholonomy dominates the strategy space.

• Pursuit–Evasion in Convex Domains: Despite an omnidirectional evader's maneuverability, proper exploitation of the DDR's kinematics (e.g., guarding, zig-zag maneuvers, and projection-tracking) guarantees capture in a bounded number of steps, with explicit bounds provided in terms of environment diameter and robot velocity (Ruiz et al., 2015).
• Surveillance Differential Games: When a DDR pursues a faster omnidirectional agent within a sensor range, the optimal feedback laws for both players partition state space into regions delineated by singular surfaces (Dispersal, Transition, Universal, Focal), dictating whether simple bang-bang or more complex singular controls are needed. The number, topology, and role of these surfaces scale with the evader-to-pursuer speed ratio (Saavedra et al., 2024).
• Assistive Control via Differential Geometry: For indirect human–robot interfacing (e.g., wheelchairs with joysticks), assistive controllers incorporating Darboux-frame kinematics provide smooth, safe trajectories, blending geometric curvature-based feedforward with direct user input, and imposing explicit speed and angle constraints (Tafrishi et al., 2022).

6. Extensions: Omnidirectional and Reconfigurable Differential Drive

Beyond the classical two-wheeled architecture, generalizations have emerged:

• Omni Differential Drive (ODD): ODD extends by using omnidirectional (Mecanum) wheels on a variable-width chassis. Modulating the lateral velocities of the collinear wheel-groups allows dynamic real-time reconfiguration of wheel base $x_2$6, enabling both omnidirectional mobility and adaptive footprint. The full kinematic mapping relates four wheel velocities to platform motion and width change, achieved without additional actuators (Zhao et al., 2024, Zhao et al., 2024).
• Skid-Steer and Tracked Variants: Unified frameworks now model skid-steer and tracked robots via the instantaneous center of rotation, bridging the gap to conventional two-wheel differential designs by extending state and control representations (Zhang et al., 2024).

7. Practical Implementation, Validation, and Future Directions

Practical realization of advanced control and estimation techniques for differential-drive robots is supported by real-world validation and hardware-centered considerations. Experimental studies range from hardware-in-the-loop object servoing via per-frame visual pose estimation (Jia et al., 2021) and robust adaptive coverage control (Razak et al., 2019), to complex navigation tasks validated through integrated EKF-based state estimation, model-based and learning-enhanced controllers, and computationally efficient optimal planning pipelines (Alwala et al., 16 Mar 2026, Li et al., 17 Nov 2025, Zhang et al., 2024).

Future work includes formal incorporation of actuator and input saturation constraints into closed-form robust control pipelines (Das et al., 5 Dec 2025), reinforcement learning for automation of gain tuning, and seamless high-level SLAM integration for mission-scale autonomy (Alwala et al., 16 Mar 2026). Advancements in scalable multi-agent planning and time-optimal path parameterization demonstrate continued progress towards deploying large heterogeneous fleets of differential-drive robots in structured and unstructured environments.