Standard control theory loves smoothness. When you push a cart, its velocity changes continuously according to forces applied. However, many of the most agile systems in the real world do not operate this way. Consider a bipedal robot walking, a quadrotor carrying a cable-suspended load that suddenly goes slack, or a wheeled-legged robot jumping over an obstacle.
These systems are governed by hybrid dynamics: they combine continuous phases of motion (such as a leg swinging through the air) with discrete, instantaneous state transitions (such as a foot striking the ground). Managing the interplay between these continuous flows and discrete resets is one of the most significant challenges in modern robotics.
The Core Intuition of Hybrid Dynamics
To understand why hybrid systems are unique, imagine modeling a simple vertical hopping robot. As demonstrated by Jorge Cham in Dynamic Stability of Open-Loop Hopping, this mechanism cannot be captured by a single set of differential equations. Instead, it must be modeled as a piecewise system that transitions through distinct operational modes: an airborne ballistic phase, a stance phase with active thrust, and a stance phase with damping and gravity.
The boundaries where these modes meet are called switching surfaces (or guards), and the sudden changes that occur when crossing them are defined by reset maps.
[ Flight Phase ] -- (Foot Touchdown / Guard) --> [ Stance Phase ]
^ |
| |
(Liftoff / Reset Map) <--------------------------------------|
In bipedal locomotion, as explained by Jessy W. Grizzle et al. in Models, feedback control, and open problems of 3D bipedal robotic walking, these transitions are governed by unilateral constraints and impulse-like forces. When a robot's foot impacts the ground, its velocity vector changes instantaneously. This creates a mathematical discontinuity: the system's state "jumps" from a pre-impact configuration to a post-impact configuration.
Why Discontinuity Breaks Traditional Control
Traditional control frameworks—like standard linear controllers or smooth trajectory optimizers—assume that a small change in the current state leads to a small change in the future state. In hybrid systems, this assumption fails catastrophically near switching surfaces.
As Alexander Winkler et al. note in An Efficient Optimal Planning and Control Framework For Quadrupedal Locomotion, the timing and nature of contact events fundamentally alter system topology. If a controller assumes continuous dynamics, a discrete transition like a foot strike acts as a massive, unmodeled disturbance.
This mismatch causes several critical issues:
- Numerical Instability: Guilhem Saurel points out in A simulation framework for simultaneous design and control of passivity based walkers that alternating between continuous motion and discrete impacts creates severe numerical sensitivities for optimization solvers, making trajectory planning highly dependent on initial guesses.
- Controller Glitches: When switching control laws abruptly between swing and stance phases, robots often experience sudden torque spikes or mechanical glitches, which require exceptionally precise state estimation to mitigate, as observed by Marco Hutter et al. in Hybrid Operational Space Control for Compliant Legged Systems.
- Invalidated Safety Guarantees: Conventional safety and stability certifications are designed for smooth motion. Frank Permenter et al. explain in Convex synthesis and verification of control-Lyapunov and barrier functions with input constraints that these limitations prevent the direct application of standard formal guarantees to walking bipeds or any system experiencing discrete contact events.
This problem extends beyond legged robots. For instance, Mrunal Sarvaiya et al. demonstrate in HPA-MPC: Hybrid Perception-Aware Nonlinear Model Predictive Control for Quadrotors with Suspended Loads that quadrotors transporting cable-suspended payloads experience severe instability if their controllers fail to model the sudden, jarring hybrid transitions between slack and taut cable states.
Propagating Stability Across Jumps
The key insight of modern hybrid dynamics research is that discrete resets do not have to be treated as destabilizing disturbances. Instead, if modeled explicitly, the discontinuities themselves can be leveraged to stabilize the system.
Hybrid Zero Dynamics and Control Lyapunov Functions
One of the most influential frameworks for managing these transitions is Hybrid Zero Dynamics (HZD). In Exponentially stable periodic walking of under-actuated biped robot, Helin Wang et al. combine HZD with an Exponentially Stabilizing Control Lyapunov Function (ES-CLF).
The core idea is to design a feedback control law that forces tracking errors to drop toward zero rapidly during the continuous phase, specifically structured so that when the system passes through an impact reset, the resulting state jump lands precisely back within a stable, lower-dimensional invariant subsystem. By coordinating the continuous swing phase with the discrete impact geometry, the robot naturally absorbs velocity jumps without losing balance.
Value Function Remapping via Reachability
To verify the safety of these systems globally, researchers calculate the Region of Attraction (RoA)—the set of all initial states from which a robot can successfully stabilize into a target gait. Jason J. Choi et al. achieve this in Computation of Regions of Attraction for Hybrid Limit Cycles Using Reachability: An Application to Walking Robots by extending Hamilton-Jacobi reachability analysis to hybrid domains.
Continuous Flow (Pre-impact) ------> [ Switching Surface ]
|
(Reset Map Remapping)
v
Continuous Flow (Post-impact) <----- Post-impact State
Instead of treating the impact surface as a boundary where analysis stops, the algorithm explicitly remaps the reachability value function from the post-impact state back to the pre-impact state at the switching surface. This mathematical remapping allows stability information to propagate backwards through the discontinuity. As a result, the controller can intentionally steer the robot into specific impacts, exploiting the physical collision to self-correct perturbations and guide the system into a stable configuration.
Representation and Learning Strategies
Because hybrid systems switch between completely different physical topologies, finding a unified mathematical representation is difficult. Recent work focuses on partitioning the state space or lifting the dynamics into higher-dimensional linear spaces to make control tractable.
Phase-Dependent Partitioning
One intuitive approach is to isolate the distinct physical regimes before attempting to learn or optimize them. In Symbolic Learning of Interpretable Reduced-Order Models for Jumping Quadruped Robots, Gioele Buriani et al. use symbolic regression to discover simple, human-readable equations for complex quadruped jumps.
They show that trying to fit a single model to an entire jump fails, but partitioning the discovery process into distinct phases based on contact states allows the system to accurately capture the discontinuous physics of the transitions.
Lifting into Linear Regimes
An alternative strategy is to eliminate the non-smoothness by transforming the coordinate system. Abulikemu Abuduweili et al. introduce this concept in Continual Learning and Lifting of Koopman Dynamics for Linear Control of Legged Robots using Incremental Koopman operators.
Nonlinear Hybrid Space (Discontinuous Impacts)
|
[ Lifted Mapping ]
v
Higher-Dimensional Latent Space (Smooth Linear Flow) ---> Standard Linear MPC
By mapping the high-dimensional, contact-driven nonlinear states into an even higher-dimensional latent space, they "lift" the hybrid dynamics into a regime where the transitions behave linearly. This allows robots to utilize highly efficient Linear Model Predictive Control (MPC) to maintain stability across sudden shifts in system physics, bypassing the need to solve complex combinatorial optimization problems in real time.
Key Foundational Work
Discrete mechanics and optimal control applied to the compass gait biped (2007) — contributed 1 concept elements to this direction.
Dynamic Stability of Open-Loop Hopping (2006) — contributed 1 concept elements to this direction.