Virtual Constraint
- Also called: virtual holonomic constraint
- Category: Control › Locomotion control
- Tags: bipedal-locomotion, nonlinear-control, hybrid-zero-dynamics, underactuation, gait-design
A relation among a robot's joint or task coordinates, such as a desired posture as a function of a phase variable, that is enforced by feedback control rather than by a mechanism.
Definition
A virtual constraint is a relationship among a robot's configuration variables that is imposed by feedback rather than by hardware. It is written as an output function y = h(q) that is zero when the constraint holds, for example the measured knee angle minus the angle the knee should have at the current point in the step. A controller drives these outputs to zero and holds them there, so the robot moves as if its joints were mechanically linked. The constraints are often parameterized by a phase variable that increases monotonically through a step, such as the angle of the stance leg or a normalized time, so the reference follows the robot's progress through the step.
The idea is central to hybrid zero dynamics, developed largely by Jessy Grizzle, Eric Westervelt and colleagues for underactuated bipedal walking. When the virtual constraints are enforced exactly, the full dynamics collapse onto a lower-dimensional surface, the zero dynamics, whose behavior can be analyzed on its own. Desired trajectories are commonly represented as Bézier polynomials of the phase variable, with coefficients found by trajectory optimization so that the resulting gait is periodic and consistent with the impacts at foot strike. Controllers enforce the constraints through input-output feedback linearization, or with control Lyapunov functions solved as quadratic programs that also respect torque limits and contact force constraints.
Outputs are not limited to joint angles. Task-space quantities such as center-of-mass position, torso orientation and swing-foot position, each minus its desired value, are common choices, especially in quadratic program controllers for quadrupeds and humanoids. The number of independent outputs cannot exceed the number of actuators, and the choice of outputs and phase variable determines how the uncontrolled degrees of freedom behave, which is why output design is a large part of using the method well.
Why it matters
Virtual constraints let engineers specify a gait as a small set of functions and enforce it with feedback, even on underactuated robots where not every degree of freedom can be controlled directly. Combined with hybrid zero dynamics, they give a way to analyze the stability of walking that includes impacts, and they underlie walking controllers on bipeds such as MABEL, ATRIAS and Cassie.
Key applications
- Underactuated bipedal walking with hybrid zero dynamics
- Task-space locomotion controllers for quadrupeds and humanoids
- Phase-based control of powered prostheses and exoskeletons
- Orbit stabilization of underactuated systems such as pendulums
- Gait libraries that switch between speeds and terrains
In practice
A team designing a gait for a planar biped chooses the stance hip angle as the phase variable and defines one output per actuated joint: each joint angle minus a Bézier polynomial of the phase. An offline trajectory optimization finds polynomial coefficients that make the gait periodic, satisfy torque and friction limits and remain consistent through the impact map at foot strike. On the robot, the controller computes the outputs and their derivatives every control cycle and solves a quadratic program for joint torques that drive the outputs to zero while keeping contact forces feasible. Because the targets depend on the phase rather than on time, a disturbance that slows the body also slows the progress of the reference, keeping the joints coordinated with the robot's actual motion.
Common misconceptions
- A virtual constraint is part of the mechanism — it is enforced only by feedback and disappears when the controller stops acting
- Virtual constraints are just time-based trajectory tracking — they are usually functions of a state-based phase variable, so the reference follows the robot's actual progress instead of a clock
- Enforcing the constraints guarantees stable walking — they fix the controlled coordinates, but the stability of the remaining zero dynamics through each impact still has to be checked
Who should know this
- Bipedal and humanoid locomotion engineers
- Prosthetics and exoskeleton control researchers
- Nonlinear control researchers working on underactuated systems
- Graduate students studying hybrid zero dynamics
Related technologies
- Hybrid Zero Dynamics
- Bézier Polynomials
- Control Lyapunov Functions
- Quadratic Programming
- Input-Output Feedback Linearization
Related terms
Related projects
- Zeroth Bot: Low-cost 3D-printed humanoid robot platform for sim-to-real reinforcement learning