For decades, roboticists and biomechanists have relied on a convenient fiction: that walking, running, and jumping can be understood by looking at a two-dimensional slice of the world. By restricting analysis to the sagittal plane (the forward-and-down plane), the complex physics of locomotion collapses into elegant, low-dimensional models. Simple templates like the Spring-Loaded Inverted Pendulum (SLIP) or reduced-order zero dynamics become mathematically tractable, allowing researchers to optimize forward velocity or vertical hopping height with ease.
However, this simplification comes at a steep cost. In the real world, locomotion is fundamentally three-dimensional. When we strip away the third dimension, we eliminate the very axis where instability is most acute: the lateral, or side-to-side, dimension. As Takeshi Mori et al. point out in their work on Reinforcement Learning for CPG-Driven Biped Robot, simplifying bipedal locomotion to a two-dimensional sagittal plane ignores the significant challenges of lateral stability and side-to-side balance. Transitioning to unconstrained three-dimensional environments introduces unmodeled degrees of freedom and complex feedback requirements that planar models simply cannot address. Without active lateral stabilization, a physical robot or biological organism will inevitably tip over sideways.
Where Planar Assumptions Break Down
The failure of planar models is not just a matter of precision; it qualitatively alters how systems behave and how we understand control. Across multiple domains, the sagittal assumption creates critical blind spots:
- Dynamic Maneuvers and Steering: A strictly planar model cannot turn or change direction. In Underactuated Motion Planning and Control for Jumping with Wheeled-Bipedal Robots, Hua Chen et al. demonstrate that while planar templates work well for linear maneuvers, they prevent robots from managing lateral stability or performing turns during dynamic movements. Similarly, the authors of Design principles for efficient, repeated jumpgliding emphasize that 2D models fail to address the roll and yaw control necessary to compensate for real-world environmental disturbances like wind gusts.
- Exoskeletons and Neurorehabilitation: Mechanical constraints in assistive devices can distort our understanding of human motor control. Leonardo Gizzi et al. note in Motor modules in robot-aided walking that many robotic gait trainers restrict movement entirely to the sagittal plane, effectively ignoring natural lateral stabilization strategies and preventing the study of multi-planar muscle synergies. This limitation hinders clinical translation; as Amy R. Wu et al. argue in An Adaptive Neuromuscular Controller for Assistive Lower-Limb Exoskeletons: A Preliminary Study on Subjects with Spinal Cord Injury, focusing exclusively on sagittal dynamics prevents the transition from tethered gait trainers to untethered, wearable exoskeletons that must independently manage multi-directional stability.
- Pathology and Motor Tracking: Clinical gait deviations and tracking errors often manifest off-axis. Christian Laßmann et al. demonstrate in Dysfunctional neuro-muscular mechanisms explain gradual gait changes in prodromal spastic paraplegia that biomechanical simulations restricted to the sagittal plane cannot capture lateral movement pathologies such as a scissoring gait or hip adductor spasticity. Furthermore, in computer vision, Marcus A Brubaker et al. show in Physics-Based Person Tracking Using Simplified Lower-Body Dynamics that abstracting 3D human motion into 2D physical dynamics ignores side-to-side swaying, forcing tracking systems to rely on unnatural kinematic heuristics that cause artifacts like "footskate."
- Upper-Limb and Postural Control: The issue extends beyond walking. Pierre Baraduc et al. show in Where Does the Population Vector of Motor Cortical Cells Point during Reaching Movements? that reaching models utilizing 2D kinematic structures fail to account for the non-linearities and increased degrees of freedom found in 3D space, underestimating the complexity of neural-to-limb transformations.
The Hidden Mechanics of Lateral Balance
Why is lateral stability so much harder to manage than sagittal stability? In forward motion, a walking system can use its forward momentum and foot placement to dynamically catch itself—falling forward is a natural part of the gait cycle. Laterally, however, the base of support is narrow, and the coupling between forward velocity and side-to-side balance is highly non-linear.
To understand how the central nervous system manages this, Enrico Chiovetto et al. studied humans balancing on a narrow beam in Low-dimensional organization of angular momentum during walking on a narrow beam. They discovered that instead of controlling individual joint angles, the brain organizes movements around regulating angular momentum. Because the medio-lateral plane represents the most unstable axis during narrow-support tasks, the body must dynamically couple side-to-side stability with forward-backward dynamics to maintain equilibrium.
When robots lack this active multi-axial coordination, they must rely on mechanical crutches. In Fast biped walking with a reflexive controller and real-time policy searching, Tao Geng et al. achieved high forward speeds using simple reflexive loops, but their controller relied on an external mechanical boom to prevent lateral instability. True autonomy requires managing roll, pitch, and yaw explicitly. As Jorge Cham demonstrates in Dynamic Stability of Open-Loop Hopping, even simple vertical hopping exhibits complex horizontal and rotational dynamics that require multi-dimensional analysis to ensure true stability.
Sagittal Plane Only (2D) Full Spatial Motion (3D)
[Forward / Down] [Multi-Axis Coupling]
│ ▲
▼ │ (Roll / Yaw / Pitch)
┌─────────────────┐ ┌────────┴────────┐
│ Linear Velocity │ │Angular Momentum │
│ Vertical Height │ │Lateral Balance │
└─────────────────┘ └────────┬────────┘
│ │
▼ ▼
(Blind to Side-to-Side (Handles Asymmetric Forces,
Perturbations & Turns) Steering, and Perturbations)
Tractable Three-Dimensional Control
If 3D models are so necessary, why did researchers avoid them for so long? The answer is computational complexity. Adding lateral and rotational degrees of freedom exponentially expands the state space, making real-time control and optimization notoriously difficult.
Modern research is overcoming this bottleneck through several distinct mathematical and algorithmic strategies:
Implicit Dynamics and Direct Collocation
Instead of explicitly solving massive differential equations at every time step, researchers are formulating system dynamics implicitly. In Efficient trajectory optimization for curved running using a 3D musculoskeletal model with implicit dynamics, Marlies Nitschke et al. used this approach to drastically reduce computation times. This mathematical formulation makes large-scale predictive simulations practical, allowing them to accurately predict the asymmetric forces of curved running using only straight-line data.
Data-Driven Linearization
In domains like autonomous driving, where weight transfer and suspension travel create complex 3D dynamics, traditional physics models are too slow for safety-critical loops. Hao Chen et al. addressed this in Deep Koopman Operator-Informed Safety Command Governor for Autonomous Vehicles by using deep learning to map non-linear 3D vehicle dynamics into a higher-dimensional linear space via the Koopman operator. This allows a linear control governor to guarantee lateral stability during aggressive maneuvers with minimal computational overhead.
Safe Zero Dynamics and Control Barrier Functions
Rather than abandoning reduced-order templates entirely, researchers are extending them to handle higher dimensions safely. The authors of Neural Gaits: Learning Bipedal Locomotion via Control Barrier Functions and Zero Dynamics Policies combine machine learning with Control Barrier Functions (CBFs). By training a neural network to satisfy set-invariance conditions on a robot's reduced-order "zero dynamics," they manage the significantly more complex balance constraints of unconstrained 3D environments, guaranteeing safety without sacrificing agility.
By embracing the full three-dimensional reality of motion—and developing the mathematical tools to handle its complexity—the fields of robotics and biomechanics are moving past the planar illusion toward truly stable, adaptive, and autonomous systems.
Key Foundational Work
Reinforcement Learning for CPG-Driven Biped Robot (2004) — contributed 1 concept elements to this direction.
Where Does the Population Vector of Motor Cortical Cells Point during Reaching Movements? (1998) — contributed 1 concept elements to this direction.