Wheeled Inverted Pendulum Model Overview

• Wheeled Inverted Pendulum Model is an underactuated, nonholonomic system that captures robot dynamics with an unstable upright equilibrium requiring feedback control.
• The model uses advanced geometric mechanics and variational integrators to derive control-affine representations for both continuous and discrete formulations.
• Practical implementations combine LQR, MPC, and online learning strategies to optimize stability, energy efficiency, and performance in high-DOF robotic systems.

A wheeled inverted pendulum model (WIPM) is a canonical, underactuated, nonholonomic mechanical system that captures the essential dynamics of robots and mechanisms that balance a rigid body atop one or two actuated wheels. These systems are characterized by an unstable upright equilibrium and strong coupling between translational and rotational degrees of freedom, requiring feedback control for stabilization. Originating in the analysis of devices such as the Segway and generalized in formal geometric mechanics, WIPMs serve as standard testbeds for control, estimation, and trajectory planning of highly nonlinear, underactuated robots. The modeling frameworks span from minimal two-DOF planar rigid-body reductions to high-DOF humanoid robots and encompass both continuous- and discrete-time, Lagrangian and symmetry-based formulations, as well as practical controller synthesis and learning-based adaptation.

1. Geometric and Dynamical Foundations

The WIPM is fundamentally an underactuated, nonholonomic system: the non-slip rolling constraint of the wheel(s) precludes arbitrary motion in the configuration space, and the net number of independent actuators is strictly less than the system’s degrees of freedom. The canonical geometric model consists of a rigid body (“body” or “pendulum”) of mass $M$ mounted on two coaxial wheels of radius $r$, separated by track width $2d$. The configuration space is typically expressed as

$$Q = (\mathbb{R}^2 \times S^1)_\text{chassis} \times (S^1 \times S^1 \times S^1)_\text{shape}$$

with coordinates $(x, y, \theta; \alpha, \phi_1, \phi_2)$: $(x, y, \theta)$ for chassis planar position and yaw, $\alpha$ for body pitch, and wheel angles $\phi_1, \phi_2$ (Gajbhiye et al., 2016).

Nonholonomic rolling constraints are expressed via one-forms: $$\omega_1 = dx - r \cos\theta (d\phi_1 + d\phi_2) = 0, \quad \omega_2 = dy - r \sin\theta (d\phi_1 + d\phi_2) = 0$$ These constraints define an Ehresmann connection, giving rise to a reduced fiber bundle structure and enabling computation of the reduced Euler–Lagrange (Lagrange–d’Alembert) equations (Gajbhiye et al., 2016).

The Lagrangian kinetic and potential energy terms for the full system typically include translational and rotational energies, along with inertial coupling terms due to CoM offset, and a gravitational restoring potential: $$L = \tfrac12(m_b+2m_W)(\dot x^2+\dot y^2) + \cdots - m_b g b \cos\alpha$$

For control-oriented analyses, specializations to planar models are common. Here, the state is $$X = \begin{bmatrix}x & \dot x & \theta & \dot\theta\end{bmatrix}^\top$$, with $x$ as the position of the wheel axle and $\theta$ as the body pitch (Zafar et al., 2018). The generalized equations of motion are: $$M(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q) = B\,\tau_w + \tau_D$$ where $\tau_w$ is the net wheel torque and $\tau_D$ lumps unmodeled disturbances.

2. Symmetry, Reduction, Nonholonomic Moments, and Connections

The gauge-theoretic perspective interprets the WIPM as a system on a principal fiber bundle, with the connection determined by nonholonomic constraints (Gajbhiye et al., 2016). The symmetry groups typically analyzed are $SE(2)$ and $SE(2)\times S^1$:

• SE(2) symmetry: full planar Euclidean symmetry, yielding “purely kinematic” reduction.
• SE(2)$\times S^1$: includes wheel phase shifts, allowing the appearance of nonholonomic momenta.

The reduced equations expose key integrals of motion (nonholonomic momenta), whose evolution encodes the irreversible momentum transfer induced by the nonholonomic coupling: $$p_1 = [ (m_b+2m_W) r^2 + 2 I_{Wyy} ]\,\dot\phi + r m_b b \cos\alpha\,\dot\alpha\ p_2 = [ I_\theta(\alpha) + (d^2/2r^2) I_{Wzz} ]\,\dot\theta$$ Their time evolution reflects energy exchange between longitudinal translation and body pitch (Gajbhiye et al., 2016).

The final compact form for feedback design includes the state vector $[\,\alpha,\,\dot\alpha,\,p_1,\,p_2\,]^\top$ and wheel torque inputs, leading to explicit, control-affine representations ready for energy shaping or tracking controllers.

3. Model Variants, Extensions, and Special Cases

Several generalizations and specializations of the canonical WIPM appear in the literature:

• Planar and multi-link reductions: High-DOF robots are often reduced to planar, single-link models for high-level control planning (e.g., by locking all joints and condensing errors into a CoM horizontal offset) (Zafar et al., 2018, Zafar et al., 2018).
• Flexible-body extensions: For robots with flexible linkages or piezo-actuated beams, state augmentation increases dimensionality and introduces coupled ODE–PDE or large-scale ODE systems, captured using modal decompositions and the extended Hamilton’s Principle to derive state-space models explicitly incorporating flexible modes and their effects on base stability (Mehrvarz et al., 2019).
• Spherical and soft-surface dynamics: Modeling of compliance at the wheel–ground interface (e.g., wheel rolling on soft surfaces) introduces set-valued differential inclusions that admit “stiction”/dead-zone friction phenomena, leading to semi-stable or chattering limit cycles when under PID control (Kiselev, 2020, Kiselev, 2020).
• Mechanical stabilization: Passive (purely mechanical) variants utilize gravity–fed braking mechanisms or coupled pendulums to achieve stabilization on slopes, subject to explicit Routh–Hurwitz stability boundaries and parametric design regions (Yoshida et al., 2015).

4. Control Architectures and Theoretical Guarantees

A wide range of feedback and optimization-based control architectures are deployed for WIPMs:

• Linear–quadratic regulation (LQR): Linearization about the upright configuration yields controllable, observable LTI systems suitable for state feedback design (Krupa et al., 2021, Detailleur et al., 21 Aug 2025, Albert et al., 2018).
• Active disturbance rejection control (ADRC): Augments LQR with an extended state observer (ESO) to estimate and cancel model disturbances—including CoM estimation errors or parametric uncertainty—enabling robust stabilization even under severely inaccurate mass models (Zafar et al., 2018).
• MPC and sparse QP solvers: Model predictive controllers exploit linearized or exact discrete-time models with predicted reference tracking and box constraints. Efficient ADMM-based sparse QP solvers demonstrated sub-10 ms solution times for real-time control on embedded hardware (Krupa et al., 2021).
• Robust/tube MPC and learning-based controllers: Robust tube-based MPC approaches are imitated by neural network controllers, whose closed-loop regional stability can be verified via sum-of-squares (SOS) Lyapunov certificates, providing provable bounds on regions of attraction and empirical RMS performance exceeding classic LQR (Detailleur et al., 21 Aug 2025).
• Hierarchical and whole-body controllers: For high-DOF humanoid WIP robots, a cascade approach is standard. High-level planners regulate the WIP zero-dynamics via simplified models (e.g., single-link reductions, CoM-based templates) and pass virtual acceleration or CoM targets to low-level QP-based inverse dynamics that enforce joint/torque/constraint limits (Zafar et al., 2018, Zafar et al., 2018, Albert et al., 2018).

5. Model Identification, Learning, and Estimation

The physical realization of the WIPM for humanoid or multi-link robots introduces high-dimensional parameterization challenges, particularly for accurate center-of-mass (CoM) tracking. Online learning frameworks condense model error into a handful of low-dimensional CoM offset parameters, updated via gradient descent from balancing postures. Meta-learning algorithms can precompute “excitory” joint poses that maximize gradient convergence, yielding rapid reduction in CoM estimation error—experimentally, mean errors decrease from several centimeters to sub-millimeter after a few hundred updates (Zafar et al., 2018).

State estimation for real hardware combines complementary filtering of inertial and encoder data, using gyroscope, accelerometer, and wheel encoder fusion to reconstruct pitch, position, and wheel velocity for controller feedback (Detailleur et al., 21 Aug 2025).

6. Discrete Mechanics, Variational Integrators, and Optimal Trajectory Planning

Control and trajectory planning over non-Euclidean configuration manifolds (e.g., SE(2) × shape) benefit from structure-preserving, variational integration. Discrete mechanics approaches discretize the system via reduced discrete Lagrangians and nonholonomic constraints, leading to state updates of the form: $$g_{k+1} = g_k \exp(-h\,A(s_k) v_k), \qquad s_{k+1} = s_k + h v_k$$ and implicit “shape” equations, enforced via a variational principle (Phogat et al., 2017, Albert et al., 2018). These integrators guarantee preservation of nonholonomic momentum maps and respect of endpoint state and control constraints within indirect optimal control synthesis (e.g., discrete-time maximum principle, dual adjoint systems).

7. Practical Implications, Performance, and Limitations

Experimental validations exhibit consistent improvement in stability metrics and energy efficiency as model accuracy and estimation converge:

• Peak power, settle-time, and maximum overshoot monotonically decrease as mass/CoM estimates improve via online learning (Zafar et al., 2018).
• On soft or compliant ground, PID control may only yield semi-stability with chattering limit cycles, constraining achievable practical performance (Kiselev, 2020).
• Discrete geometric integration and whole-body QP controllers are essential for scaling to high-DOF, high-performance balancing tasks with simultaneous posture and mobility objectives (Zafar et al., 2018, Albert et al., 2018).

Foundational limitations include:

• Reduced planar models do not capture full 3D or non-planar dynamics.
• Online gradient descent only learns a CoM-equivalent, not unique mass/inertia parameters (Zafar et al., 2018).
• Model linearizations are valid only near fixed postures; large deviations require nonlinear or “relinearize at each step” approaches.
• Robustness to modeling errors relies on observer bandwidth and the correct characterization of disturbance torques.

---

References

• (Zafar et al., 2018) S. Gajbhiye et al., "Online Center of Mass Estimation for a Humanoid Wheeled Inverted Pendulum Robot"
• (Gajbhiye et al., 2016) S. Gajbhiye et al., "Symmetries in the wheeled inverted pendulum mechanism"
• (Zafar et al., 2018) K. Sreenath et al., "Hierarchical Optimization for Whole-Body Control of Wheeled Inverted Pendulum Humanoids"
• (Kiselev, 2020) A. Kiselev, "Stabilization of the wheeled inverted pendulum on a soft surface"
• (Mehrvarz et al., 2019) A. Mehrvarz et al., "A New Dynamic Model of a Two-Wheeled Two-Flexible-Beam Inverted Pendulum Robot"
• (Krupa et al., 2021) D. Braun et al., "Real-time implementation of MPC for tracking in embedded systems: Application to a two-wheeled inverted pendulum"
• (Detailleur et al., 21 Aug 2025) C. Rainwater et al., "Synthesis and SOS-based Stability Verification of a Neural-Network-Based Controller for a Two-wheeled Inverted Pendulum"
• (Phogat et al., 2017) R. Phogat et al., "Structure-preserving discrete-time optimal maneuvers of a wheeled inverted pendulum"
• (Yoshida et al., 2015) K. Yoshida et al., "Nonlinear analysis on purely mechanical stabilization of a wheeled inverted pendulum on a slope"
• (Albert et al., 2018) R. Banavar et al., "Structure-Preserving Constrained Optimal Trajectory Planning of a Wheeled Inverted Pendulum"