Differential-Drive Robot

• Differential-drive robots are mobile platforms with two independently actuated, coaxial wheels that enable planar translation and rotation.
• Research focuses on precise kinematic and dynamic modeling, robust control under uncertainty, and real-time trajectory optimization.
• Advanced implementations are validated through hardware experiments and extend to multi-agent coordination and omnidirectional control scenarios.

A differential-drive robot (DDR) is a class of nonholonomic mobile robot characterized by two independently actuated coaxial wheels, optionally augmented by passive caster(s) for stability. The canonical DDR executes planar motion by modulating the speeds of its left and right drive wheels, effecting both translation and rotation. This architecture, central to both theoretical algorithm development and broad application domains, underlies a wide spectrum of contemporary mobile robot research, encompassing robust control under uncertainty, multi-agent planning, learning-based tracking, geometric assistance for human operators, and extensible hardware variants such as omnidirectional and reconfigurable drives.

1. Kinematic and Dynamic Modeling

The foundational state space for a DDR is $\xi = [x, y, \theta]^\top$, with controls given as forward velocity $v$ and angular velocity $\omega$ (Das et al., 5 Dec 2025, Ali et al., 2024, Ruiz et al., 2015). The standard kinematic equations are

$$\dot{x}=v\cos\theta,\ \dot{y}=v\sin\theta,\ \dot{\theta}=\omega$$

relating wheel velocities via

$$v = \frac{r}{2}(\omega_r+\omega_l),\quad \omega = \frac{r}{d}(\omega_r-\omega_l),$$

where $r$ is wheel radius and $d$ is half the track width (Ali et al., 2024). Higher-fidelity models incorporate mass $m$, inertia, wheel friction (viscous and Coulomb), actuator dynamics, and sometimes a COM offset, yielding state vectors of up to 10 dimensions and dynamic equations with strong coupling and nonlinearities (Pishkhani, 30 Aug 2025, Alwala et al., 16 Mar 2026).

Uncertainties—arising from modeling errors or external disturbances—are modeled as additive, bounded, and (often) locally Lipschitz disturbances $d(\xi, t)$ (Das et al., 5 Dec 2025). Discrete-time models are widely used for controller synthesis and account for sampling and system/measurement noise, typically via zero-order-hold discretization (Nguyen et al., 2020).

2. Optimal Control and Robustness under Uncertainty

DDR control design must respect the nonholonomic constraint: DDRs cannot move laterally, necessitating specialized planning and feedback strategies. Classical approaches include decoupling-based PID, feedback linearization, computed-torque control, and Lyapunov-based switched controllers (Pishkhani, 30 Aug 2025, Nguyen et al., 2020, Kouvakas et al., 7 Oct 2025).

Handling uncertainty and disturbances, several directions have emerged:

• Spatiotemporal Tube (STT)-based control constructs a time-varying safe corridor $$\Gamma(t) = \{x\in\mathbb{R}^2 : \|x-c(t)\| \leq r(t)\}$$ and provides a formally verified, sampling-based synthesis algorithm for tube construction, ensuring temporal reach-avoid-stay (T-RAS) constraints are met in the presence of bounded disturbances (Das et al., 5 Dec 2025). Analytic, approximation-free closed-form control guarantees forward invariance within the STT.
• Adaptive and learning-based control leverages Radial Basis Function Neural Networks (RBF-NN) to compensate unknown disturbance dynamics online. The RBF-NN output augments a feedback-linearizing control law, achieving provably stable and robust trajectory tracking in concert with multi-modal EKF state estimation, reducing velocity errors up to 53.9% in real-world experiments (Alwala et al., 16 Mar 2026).
• Robust Lyapunov control in the presence of noise employs regionally switched feedback, dividing the task into "global" and "local" configurations. By tuning discrete Lyapunov functions with respect to measurement and system noise bounds, stability and rapid convergence to the goal are retained even under realistic sensor and actuation uncertainties (Nguyen et al., 2020).

3. Planning, Trajectory Optimization, and Multi-Agent Coordination

DDR motion planning must address nonholonomic, dynamic, and environmental constraints.

• Trajectory optimization frameworks compatible with all differential-drive variants (two-wheeled, skid-steering, tracked) parameterize motion using polynomials of angular and linear velocity or their integrals. Optimization problems minimize trajectory roughness under kinematic and safety constraints, and are solved efficiently (in milliseconds) using L-BFGS or SOCP, with full-stack systems (including EKF-based slip estimation and NMPC tracking) delivering real-time, robust performance in unstructured environments (Zhang et al., 2024, Li et al., 17 Nov 2025).
• Time-Optimal Path Parameterization (TOPP-DWR) incorporates linear, angular, and joint (wheel) velocity as well as acceleration constraints using piecewise-constant angular velocity over non-uniform B-spline paths. The parameterization leads to a convex SOCP, tractable at high discretization and validated to enforce constraints in both simulation and field robotics (Li et al., 17 Nov 2025).
• Multi-agent motion planning frameworks (e.g., MASS) employ a bi-level hierarchy: high-level MAPF for collision-free logical task assignment and a mid-level stationary-state planner for kinodynamically feasible DDR execution. A low-level solver generates dynamically admissible speed profiles subject to speed/acceleration bounds, allowing plans for up to 150 agents with empirically demonstrated throughput gains up to 400% in congested warehouse environments (Yan et al., 2024).

4. Hybrid and Intelligent Control Approaches

Several recent works integrate model-based and data-driven techniques for superior performance:

• Gray-box Computed Torque Control replaces a DRL policy with a computed torque controller whose parameters are optimized via deep reinforcement learning (e.g., TD3). Key parameters such as pole locations, friction coefficients, and dynamic constants are bounded and learned, enabling critically damped responses and high tracking performance with few learning episodes—bridging classical and data-driven control (Pishkhani, 30 Aug 2025).
• Model matching and delay compensation address challenges in teleoperated robots under network-induced delays via multi-layered nonlinear feedback structures and dynamic precompensators. Closed-form analytic synthesis ensures exact model matching of heading commands and stability up to computable delay bounds (Kouvakas et al., 7 Oct 2025).
• Geometric and assistive controllers employ Darboux-frame geometric formalism and joystick blending to assist human operators of DDRs (such as electric wheelchairs) in real-time, producing smooth and safe trajectories without specifying explicit desired states (Tafrishi et al., 2022).
• Control barrier functions (CBF) with dynamic feedback linearization are embedded in MPC formulations for guaranteeing DDR collision avoidance in real-time (Ali et al., 2024). DFL transforms the nonlinear, underactuated system into a linear chain of integrators for efficient MPC with barrier constraints.

5. Extensions: Hardware Variants and New Paradigms

Recent research has generalized the classical DDR:

• Omni Differential Drive (ODD) introduces two variable-width, collinear omnidirectional wheel-groups enabling simultaneous control of $v$0, yaw, and inter-group separation $v$1. Lateral differential drive (LDD) exploits side-slip capabilities to reconfigure footprint on-the-fly without additional actuators, retaining the DDR's simplicity but providing holonomic mobility and significant adaptability in cluttered or dynamic environments (Zhao et al., 2024, Zhao et al., 2024).
• DDR in pursuit-evasion and surveillance differential games: Theoretical results provide time-optimal strategies for capturing faster omnidirectional evaders or sustaining surveillance, revealing a mosaic of region-dependent bang-bang and singular controls dictated by the structure of the value function and speed ratios. Four classes of singular surfaces define the system's regime-switching logic (Saavedra et al., 2024, Ruiz et al., 2015).
• Stochastic and temporal logic-constrained DDR control: Formal policy synthesis tools map the uncertain, noisy DDR dynamics to Markov Decision Processes and employ model checking to generate feedback that maximizes probability of meeting rich, time-bounded temporal logic specifications, with provable lower-bounds on satisfaction probability validated in hardware (Cizelj et al., 2012).

6. Experimental Validation and Practical Considerations

DDR research involves rigorous hardware-based experimentation and realistic performance metrics:

• Robust state estimation leverages multi-sensor fusion with EKF, integrating IMU, wheel odometry, visual, and LiDAR cues (Alwala et al., 16 Mar 2026).
• Kinematic and dynamic parameter identification is deployed online via EKF and fused into full-stack trajectory optimization and tracking pipelines (Zhang et al., 2024).
• Hardware platforms range from prototypical TRACER, SCOUT, and CubeTrack robots, equipped with varied sensor suites, to commercial electric wheelchairs and self-balancing platforms implementing advanced drive and control architectures (Zhao et al., 2024, Tafrishi et al., 2022, Pishkhani, 30 Aug 2025).
• Sample-efficient learning controllers have been demonstrated in simulation (MuJoCo) and on physical robots, with convergent behavior and clear advantages in tracking error and settling times (Pishkhani, 30 Aug 2025).
• Scalability is established in multi-agent domains, e.g., hundreds of DDRs executing lifelong warehouse rearrangement tasks with throughput superior to prior approaches (Yan et al., 2024).

7. Open Problems and Future Directions

Ongoing research seeks to:

• Integrate actuator limits and explicit input constraints into verified tube-based or optimal control synthesis (Das et al., 5 Dec 2025).
• Extend control-theoretic guarantees to higher-order, underactuated, or morphologically reconfigurable differential-drive platforms (Zhao et al., 2024, Zhao et al., 2024).
• Enhance learning-based adaptation via reinforcement learning or hierarchical planners for joint perception and control (Alwala et al., 16 Mar 2026, Pishkhani, 30 Aug 2025).
• Formalize scalable, real-time, and provably safe multi-agent planning frameworks under uncertainty and cooperative/competitive objectives (Yan et al., 2024, Saavedra et al., 2024).
• Advance human-in-the-loop and geometry-based differential-drive control for assistive and autonomous mobility solutions in unstructured settings (Tafrishi et al., 2022).

The DDR remains a foundational architecture for robotics research due to its rich theoretical structure, strong practical utility, and enduring potential for innovation at the intersection of modeling, planning, control, and learning.