Reduced Order Models for Real Time Locomotion Planning

138 concepts107 papers1 proposal

When controlling a high-degree-of-freedom robot—such as a humanoid with dozens of joints or a quadruped navigating rough terrain—generating stable, dynamic motions in real time is a monumental challenge. If a trajectory planner attempts to optimize every joint angle, motor torque, and contact force simultaneously, the underlying mathematical problem becomes highly non-linear, non-convex, and computationally intractable for real-time control loops.

To bypass this computational bottleneck, modern robotics relies on a powerful insight: we do not need to model the entire robot to plan its global motion. Instead, we can project the complex, high-dimensional physics of the robot onto a simplified, low-dimensional "template model." This approach, known as reduced-order dynamic modeling, captures the essential physics required for balance and locomotion while discarding the computational baggage of full-body dynamics.

The Core Intuition: Centroidal Dynamics

No matter how complex a robot's internal mechanics are, its interaction with the environment is governed entirely by external forces—primarily gravity and ground reaction forces. The net effect of these external forces dictates the motion of the robot’s total Center of Mass (CoM) and its net angular momentum. This behavior is captured by centroidal dynamics.

As highlighted by Scott Kuindersma et al. in Optimization-based locomotion planning, estimation, and control design for the atlas humanoid robot, focusing on centroidal momentum provides a sufficient abstraction for planning highly dynamic maneuvers like jumping or running. By treating the robot's total linear and angular momentum as the primary control targets rather than individual joint inertias, planners can ensure physical feasibility without getting bogged down by the robot's specific kinematic configuration.

This abstraction drastically simplifies the optimization landscape. For instance, in human motion capture and animation, Davis Rempe et al. demonstrate in Contact and Human Dynamics from Monocular Video that optimizing a reduced-dimensional centroidal model is enough to eliminate physical artifacts like foot-skating, bypassing the need for complex full-body muscle or joint torque simulations.

A Spectrum of Template Models

Depending on the robot's morphology and the target maneuver, researchers choose from a variety of reduced-order template models that balance simplicity with physical fidelity.

Inverted Pendulums

The simplest templates treat the robot's mass as concentrated at a single point. The classic Linear Inverted Pendulum Model (LIPM) assumes the CoM moves along a fixed horizontal plane, reducing the dynamics to linear equations. Alexander Winkler et al. show in Fast Trajectory Optimization for Legged Robots using Vertex-based ZMP Constraints that such template models can effectively generate complex walking, trotting, and recovery behaviors for high-degree-of-freedom hardware when paired with high-frequency low-level tracking controllers.

For more agile or hybrid robots, the pendulum model can be extended:

  • Variable-Length Wheeled Inverted Pendulum (VL-WIP): Designed for wheeled-legged robots, this model abstracts the system into a prismatic joint and a wheel. As shown by Traiko Dinev et al. in Modeling and Control of a Hybrid Wheeled Jumping Robot, this abstraction successfully captures CoM height adjustments and wheel-ground interactions, allowing the robot to discover complex behaviors like jumping over gaps from first principles.
  • Wheeled-Spring-Loaded Inverted Pendulum (W-SLIP): To model explosive jumping maneuvers in wheeled bipeds, Hua Chen et al. introduce a template in Underactuated Motion Planning and Control for Jumping with Wheeled-Bipedal Robots that incorporates wheel mass and rotational dynamics, capturing the critical underactuated coupling between rolling momentum and vertical leg thrust.

Single Rigid-Body Models (SRBM)

When a maneuver involves significant rotations—such as backflips or high-speed cornering—point-mass models fall short because they neglect rotational inertia. The Single Rigid-Body Model steps up the fidelity by treating the entire robot as a single rigid mass with a fixed or changing inertia tensor.

In Optimizing Bipedal Maneuvers of Single Rigid-Body Models for Reinforcement Learning, Ryan Batke et al. leverage an SRBM to generate expert reference trajectories. This abstraction captures essential 3D orientation and angular momentum physics, accelerating reinforcement learning training efficiency by 2.5x and enabling hardware to execute high-speed bipedal runs.

Multi-Mass Approximations

When the mass of the limbs themselves significantly impacts balance (e.g., during a heavy kick), a single rigid body may be insufficient. Intermediate models bridge this gap by distributing mass across a few discrete points. For example, Grzegorz Ficht and Sven Behnke develop an analytical whole-body control framework in Fast Whole-Body Motion Control of Humanoid Robots with Inertia Constraints using a five-mass model (representing the limbs and trunk). This reduction allows the system to approximate complex rotational inertia variations in tens of microseconds.

Hierarchical Control: Bridging Template to Hardware

How does a trajectory planned on a simplified template actually move a physical robot with dozens of joints? The answer lies in a hierarchical control architecture.

┌────────────────────────────────────────────────────────┐
│               High-Level Global Planner                │
│  Uses simplified kinematic/point-mass models to find   │
│  feasible paths through environments.                  │
└───────────────────────────┬────────────────────────────┘
                            │ (CoM path, footholds)
                            ▼
┌────────────────────────────────────────────────────────┐
│          Mid-Level Trajectory Optimization             │
│  Solves Centroidal Dynamics / SRBM / Template MPC      │
│  to compute optimal ground reaction forces.            │
└───────────────────────────┬────────────────────────────┘
                            │ (Target forces, momentum)
                            ▼
┌────────────────────────────────────────────────────────┐
│           Low-Level Whole-Body Controller              │
│  Resolves full multi-body kinematics and torques via   │
│  high-frequency Quadratic Programming (QP).            │
└────────────────────────────────────────────────────────┘
  1. High-Level Planning: A global planner uses highly abstracted models (such as point masses or simplified kinematics) to find navigable routes through complex environments. For instance, Hao Luan and Ziqi Zhao demonstrate in Collaborative Trolley Transportation System with Autonomous Nonholonomic Robots that abstracting a multi-body system into a simplified kinematic model lowers computational overhead enough to allow real-time navigation through constrained spaces.
  2. Mid-Level Optimization: A Model Predictive Control (MPC) loop uses a template model (like an SRBM or centroidal model) to compute the optimal trajectories for the center of mass and the required ground reaction forces over a short time horizon.
  3. Low-Level Execution: A high-frequency whole-body controller (often running at 1 kHz) ingests the target centroidal forces and maps them down to individual joint torques using full-body inverse dynamics. This low-level loop absorbs the residual errors and unmodeled dynamics introduced by the template simplification.

By decoupling global physics from local joint coordination, reduced-order modeling makes real-time, reactive, and highly dynamic robotic locomotion possible on physical hardware.

Key Foundational Work

Walking Control And Dynamics Of A Systems With Two Legs (1975) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_reduced_order_dynamic_modeling_0d38ed3c,
  title = {reduced-order dynamic modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/reduced-order-dynamic-modeling/rdr_0d38ed3caee740e4b3a15de4ac784b31},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_reduced_order_dynamic_modeling_0d38ed3c_v1,
  title = {reduced-order dynamic modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/reduced-order-dynamic-modeling/rdr_0d38ed3caee740e4b3a15de4ac784b31/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake