Reduced-Order Model (ROM)

The primary function of a ROM is to drastically reduce the number of state variables from the high-dimensional full-order model (FOM). A bipedal robot's FOM may have dozens of degrees of freedom (joint angles, velocities). A ROM like the Linear Inverted Pendulum Model (LIPM) reduces this to just a few key states: the Center of Mass (CoM) position and velocity. This reduction makes complex optimization and real-time control computationally feasible by focusing only on the dynamics that dominate the task, such as balance during walking.

ROMs are designed to have closed-form or easily computable solutions. The LIPM dynamics, for instance, are linear and can be integrated forward in time with simple equations. This allows for:

• Instantaneous prediction of future CoM motion.
• Exact calculation of stability metrics like the Capture Point.
• Fast optimization for footstep planning over many steps. This tractability is critical for generating reactive, real-time control policies in unpredictable environments, where solving the full non-linear dynamics would be prohibitively slow.

A ROM is not a universal model; it is tailored to a specific locomotor behavior by making deliberate simplifying assumptions.

• LIPM for Walking: Assumes constant CoM height and no angular momentum, perfectly capturing the sagittal/lateral balance problem.
• Spring-Loaded Inverted Pendulum (SLIP) for Running: Models the leg as a massless spring, abstracting away leg mass to capture the essential energy exchange of hopping and running.
• Single Rigid Body Dynamics for Agile Motion: Assumes the robot is a single rigid body with variable angular momentum, useful for dynamic maneuvers like jumping. The art of ROM design lies in choosing which complexities to ignore without losing predictive power for the target behavior.

ROMs operate at an intermediate level of abstraction, connecting high-level intent to low-level actuation.

1. High-Level Planner: Uses the ROM (e.g., LIPM) to plan a sequence of stable foot placements and CoM trajectories over several seconds.
2. Mid-Level Controller: A Model Predictive Controller (MPC) uses the same ROM to compute optimal Ground Reaction Forces (GRFs) for the immediate future (~0.5 seconds) to track the planned trajectory.
3. Low-Level Controller: A Whole-Body Controller (WBC) uses the desired GRFs from the MPC, along with the full robot dynamics, to solve for the actual joint torques that realize those forces. The ROM provides a consistent, simplified dynamic language that each layer can use efficiently.

The simplifying assumptions of a ROM create a reality gap between its predictions and the actual robot's behavior. Key discrepancies include:

• Unmodeled dynamics (e.g., leg swing inertia, joint flexibility).
• Disturbances (e.g., uneven ground, pushes).
• Actuation limits and delays. Successful deployment requires compensation strategies:
• Robust Control: Designing controllers that are insensitive to model errors.
• Online Adaptation: Using state estimation to update the ROM's parameters (like effective CoM height) in real-time.
• Hierarchical Stability: Relying on the low-level WBC to enforce contact constraints and manage details the ROM ignores.

This conceptual framework formalizes the role of ROMs in locomotion control.

• Template: The ROM itself (e.g., SLIP model). It defines the target dynamics—the simple, desired behavior.
• Anchor: The physical, high-dimensional robot. It is the system being controlled. The control objective is to anchor the full robot's behavior to the template dynamics. The WBC and MPC act as the 'glue' that maps the template's solutions (desired forces, motions) into feasible commands for the anchor. This separation of concerns is fundamental to managing complexity; the template handles the 'what' (the gait objective), and the anchor's controllers handle the 'how' (the physical realization).

The Linear Inverted Pendulum (LIP) model is the cornerstone for bipedal walking analysis. It simplifies the robot to a point mass atop a massless, telescoping leg, with two key assumptions: the center of mass moves at a constant height, and the dynamics are linearized. This creates a divergent component of motion, making foot placement critical for stability. It is the basis for the Zero-Moment Point (ZMP) criterion and Capture Point theory, enabling real-time Model Predictive Control (MPC) for humanoid robots like Boston Dynamics' Atlas.

The Spring-Loaded Inverted Pendulum (SLIP) model is the canonical template for running, hopping, and dynamic gaits like the trot. It represents a leg as a massless, linear spring attached to a point mass body. The model captures the passive, energy-conserving dynamics of the stance phase, where kinetic energy is stored as elastic potential energy in the spring and then returned for propulsion. It explains the self-stabilizing properties of animal running and is used to design compliant controllers and reactive locomotion strategies for quadrupeds and bipeds.

The Angular Momentum Pendulum Model extends the LIP by accounting for the robot's centroidal angular momentum. In the full LIP, the centroidal moment is assumed zero. This model incorporates a flywheel or reaction wheel abstraction at the center of mass, allowing the controller to generate and manage angular momentum. This is essential for executing dynamic maneuvers like jumping, turning, and recovering from large pushes, where swinging limbs or a torso must be used to regulate rotational dynamics while keeping feet stationary.

The 3D LIP with Foot Rotation model introduces a foot with a finite size and the ability to rotate about its edges (heel and toe). This relaxes the point-contact assumption of the basic LIP, allowing the Center of Pressure (CoP) to move within the foot. The model is crucial for planning step-to-step transitions, toe-off and heel-strike behaviors, and for achieving more natural, human-like walking with underactuated ankle joints. It provides a bridge between simple template models and the full floating base dynamics of a humanoid.

The Rimless Wheel and 2D Compass Gait models are the simplest templates for studying passive dynamic walking. The Rimless Wheel is a polygonal wheel that rolls down a shallow slope under gravity, demonstrating limit-cycle stability. The Compass Gait is a two-legged, point-mass walker with no knees or torso. These models prove that stable, periodic walking can emerge from purely mechanical dynamics with no actuation or control, inspiring the design of energy-efficient, underactuated robots and informing analyses of gait stability and basins of attraction.

The Variable Height Linear Inverted Pendulum model generalizes the LIP by allowing the center of mass height to change. This is critical for motions like crouching, stair climbing, or energy-modulating running. Controlling the height directly influences the natural frequency of the pendulum. This model is often managed through the Divergent Component of Motion (DCM) and its time-derivative, the Virtual Repellent Point (VRP). By planning trajectories for these reduced-order states, controllers can generate complex 3D motions while maintaining the stability guarantees of the simpler LIP framework.

The Linear Inverted Pendulum Model (LIPM) is the quintessential Reduced-Order Model for bipedal walking. It simplifies a robot to a point mass atop a massless telescoping leg, assuming a constant center of mass height. This linearization yields closed-form solutions for dynamics, making it the basis for real-time footstep planning and balance controllers like those used in Honda's ASIMO. Its key assumptions enable the derivation of critical stability metrics, including the Capture Point and Divergent Component of Motion (DCM).

The Spring-Loaded Inverted Pendulum (SLIP) model is a template ROM for running, hopping, and dynamic gaits. It represents a leg as a massless spring attached to a point mass body. This model captures the passive, energy-conserving dynamics of stance phase, where kinetic energy is stored as elastic potential energy in the "spring" and returned for propulsion. It is fundamental to understanding energy-efficient locomotion and informs the design of compliant actuators and control policies for robots like the MIT Cheetah.

The Capture Point is a stability criterion derived directly from the LIPM ROM. It defines the point on the ground where a robot must place its foot to bring its center of mass to a complete stop atop that foot in a single step. Mathematically, it is the sum of the current center of mass position and its velocity scaled by the pendulum frequency. This concept transforms complex balance recovery into a geometric foot placement planning problem and is central to reactive push recovery strategies.

The Divergent Component of Motion (DCM) is a state variable extracted from the LIPM ROM that encapsulates the unstable part of the center of mass dynamics. It is defined as ξ = x + (ẋ / ω), where x is the CoM position and ω is the natural frequency of the pendulum. Controlling the DCM's trajectory (often to a predefined DCM offset from the foot) ensures the overall system's stability. It provides a more direct method for trajectory generation than raw CoM planning.

Model Predictive Control (MPC) is an advanced control methodology that heavily relies on a dynamic model—often a ROM—to optimize future robot behavior. In legged locomotion, an MPC controller uses the LIPM or similar ROM to predict the system's state over a finite horizon. It then solves an optimization problem (e.g., minimizing tracking error or energy) at each control cycle to determine optimal ground reaction forces and footstep locations before applying the first step of the plan. This allows for proactive adaptation to terrain.

Whole-Body Control (WBC) is a hierarchical control framework that uses the full, high-dimensional robot model. It typically sits downstream of a planning layer that uses a ROM. The WBC receives desired tasks from the planner—like desired foot forces from an LIPM-based MPC—and solves a Quadratic Program (QP) to compute joint torques that realize these tasks while satisfying all physical constraints (torque limits, friction cones, kinematics). This bridges the simplicity of the ROM with the physical reality of the complex robot.