Quadratic Program Controllers

• Quadratic Program controllers are optimization-based feedback mechanisms that solve convex quadratic problems at every control step.
• They integrate system dynamics, physical limits, and safety constraints in real-time to ensure robust performance in robotics and autonomous systems.
• Hierarchical architectures and accelerated solvers enable these controllers to achieve high-frequency stability and coordinated multi-agent control.

Quadratic Program (QP) Controllers

Quadratic Program (QP) controllers constitute a class of real-time optimization-based feedback controllers in which each control action is computed as the solution to a convex quadratic program, typically at every discrete control step. This paradigm enables systematic handling of system dynamics, safety-critical constraints (e.g., friction cones, collision avoidance), input and state limits, and multi-objective tradeoffs via quadratic costs, all within a computationally tractable convex optimization framework. QP controllers are pivotal in robotics, autonomous systems, and safety-critical control, supporting hierarchical and structured stabilization, safety, and coordinated multi-agent behaviors under complex physical and task constraints.

1. Core Structure and Mathematical Formulation

At their core, QP controllers determine the current control input $u^*$ by solving an optimization problem of the form

$$\min_{u\in \mathbb{R}^m} \quad \frac{1}{2} u^\top H u + F^\top u,$$

subject to a collection of affine inequality (and possibly equality) constraints,

$A^\top u \leq B,$

where $H \succ 0$ enforces strict convexity, $F$ encodes objectives such as reference tracking, and $(A, B)$ collect all system, safety, or operational constraints relevant at the current control timestep (Pitroda et al., 2024, Garg et al., 2019, Wu et al., 2022). The QP can be extended to include slack variables for constraint softening, auxiliary states, or temporally stacked decision variables when implementing preview or receding-horizon schemes.

This structure is highly general; it subsumes:

• Real-time whole-body control with explicit dynamic constraints and friction cones (Pitroda et al., 2024, Kuindersma et al., 2013, Hamed et al., 2020).
• Safety-critical control via Control Barrier Functions (CBFs) and/or Control Lyapunov Functions (CLFs) recast as affine constraints (Garg et al., 2019, Reis et al., 2024, Miao et al., 23 Apr 2025).
• Hierarchical or multi-domain control with QP layers for reduced-order models, followed by whole-body inverse dynamics (Pitroda et al., 2024, Hamed et al., 2020).
• Multi-agent and cooperative transportation where QP outputs (e.g., contact forces, reference velocities) enter as virtual commands for subordinate tracking layers (Wu et al., 14 Dec 2025, Wu et al., 2022).

2. Controller Synthesis and Constraint Typology

2.1 Physics-Based Constraints

QP controllers natively represent fundamental physical limits by formulating them as affine or polyhedral inequalities:

• Friction Cones and Ground Reaction Constraints: Linearized forms $|\lambda_x| \leq \mu\lambda_z$, $\lambda_z \geq 0$ enforce contact unilaterality and friction cones in legged or manipulation tasks (Pitroda et al., 2024, Kuindersma et al., 2013, Hamed et al., 2020).
• Torque, Actuator, and Joint Limits: Direct box or polyhedral constraints $\tau \in [\tau_{\min}, \tau_{\max}]$ (Kuindersma et al., 2013, Djeha et al., 2023).
• Contact/Force Closure and Feasibility: For cooperative transport, the space of feasible contact forces is enforced via constraints on virtual agent inputs (Wu et al., 14 Dec 2025).

2.2 Stability and Safety Certificates

• Control Lyapunov Functions (CLFs): Stability objectives are encoded as affine (continuous-time) or quadratic (sampled-data) constraints based on the Lyapunov decrease rate. Sampled-data stabilization uses Quadratically Constrained QPs (QCQPs) (Taylor et al., 2021).
• Control Barrier Functions (CBFs): Forward invariance of safety sets is ensured by expressing the infinitesimal decrease of safety functions as affine constraints, e.g., $L_f h + L_g h u + \gamma h \geq 0$ (Garg et al., 2019, Miao et al., 23 Apr 2025, Reis et al., 2024).
• Multi-Objective Patchwork: Modern frameworks simultaneously handle multiple CBFs and CLFs by stacking their constraints (Reis et al., 2024, Garg et al., 2019).

2.3 Spatio-Temporal and Task Constraints

• Prescribed-Time and Spatio-Temporal Specifications: QPs synthesize reachable sets and temporal logic specifications by imposing constraints that ensure reachability within a user-defined time horizon (Garg et al., 2019, Garg et al., 2019, Black et al., 2020).
• Stage Switching and Multi-Segment Coordination: Temporal constraint stacking enables piecewise objectives with guaranteed transitions and joint forward invariance (Garg et al., 2019).

3. Hierarchical and Modular QP Architectures

QP controllers are routinely integrated into hierarchical multi-resolution feedback schemes, decoupling modeling complexity and computational load:

• Reduced-Order Trajectory Optimization: A high-level QP solves for virtual variables (e.g., contact forces, CoM trajectories) under simplified models (e.g., VLIP, ZMP) (Pitroda et al., 2024, Kuindersma et al., 2013, Hamed et al., 2020).
• Whole-Body Mapping: The QP solution feeds into a high-frequency inverse-dynamics loop, distributing contact forces and reference accelerations across the robot’s full degrees of freedom (Pitroda et al., 2024, Hamed et al., 2020, Djeha et al., 2023).
• Real-Time Feasibility and Constraint Selection: LP-based pre-solve routines assess QP feasibility and enable dynamic constraint activation/deactivation for robust operation under shifting task constraints (Rousseas et al., 17 Feb 2025).

Table: Role of QP Phases in a Bipedal Locomotion Controller (Pitroda et al., 2024)

QP phase	Task	Solve Rate
Stance-Leg QP	Sagittal GRFs/tracking	100 Hz
Whole-Body Mapper	Distribute torques/thrusters	2 kHz

4. Theoretical Guarantees and Regularity

QP controllers inherit well-understood properties from convex optimization:

• Continuity and Uniqueness: Solutions are unique and locally Lipschitz in the problem parameters and current state under strict convexity and regularity of active sets (Garg et al., 2019, Wu et al., 2022, Agrawal et al., 25 Aug 2025, Wu et al., 14 Dec 2025).
• Global Feasibility: Slack variable augmentation, positive-basis constraint construction, and feasible-set reshaping guarantee robust feasibility even under constraint redundancy or actuation uncertainties (Wu et al., 2022, Rousseas et al., 17 Feb 2025).

Recent work recasts QP-based controller maps into globally Lipschitz closed-form via Second-Order Cone Program (SOCP) reformulation, circumventing dependence on classical constraint qualifications such as LICQ (Linear Independence Constraint Qualification) (Agrawal et al., 25 Aug 2025).

• Closed-Loop Safety and Stabilization: Theoretical analyses provide conditions under which QP-based controllers guarantee forward invariance (Nagumo's theorem), set-stability, and, via small-gain interconnection, overall convergence in interconnected multi-system settings (Garg et al., 2019, Wu et al., 2022, Wu et al., 14 Dec 2025).
• Undesirable Equilibria: Multi-CBF QP frameworks can introduce undesired closed-loop equilibria at CBF boundaries, with their local stability determined by curvature relations between CLF and CBFs; compatibility and QP-based CLF reshaping strategies have been advanced for quasi-global convergence with safety (Reis et al., 2024).

5. Computational Methods, Performance, and Implementation

QP controllers have achieved high-throughput real-time operation via fast custom or general-purpose solvers, GPU acceleration, and tailored active-set methods:

• Solver Accelerations: Active-set solvers leveraging problem sparsity and warm-starts enable solve times on the order of 0.2–0.5 ms for full-body controllers (1 kHz) (Kuindersma et al., 2013, Pitroda et al., 2024, Hamed et al., 2020). GPU-accelerated solver architectures, such as ReLU-QP, directly map iterative algorithms (ADMM) to deep ReLU networks executable on machine learning hardware, delivering order-of-magnitude speed-ups for large MPC problems (Bishop et al., 2023).
• Hybrid Algorithms: Operator-splitting and Newton-hybrid methods such as Newton-PIPG provide global convergence and local quadratic acceleration for optimal-control QPs with structured constraints, exploiting problem block-tridiagonality for linear-time factorization (Luo et al., 28 Mar 2025).
• Real-Time Feasibility Check: Duality-based LP formulations allow efficient infeasibility detection and constraint set adaptation, crucial for online control (Rousseas et al., 17 Feb 2025).

Table: Example Solver Performance for Whole-Body Walking QP (Kuindersma et al., 2013)

Solver	Average Solve Time (ms)	Control Rate
Custom Active-Set	0.20	1 kHz
Gurobi (dual simplex)	1.0	1 kHz
CVXGEN	2.2	–

6. Applications and Extensions

QP controllers have been key enablers in advanced robotics and control tasks:

• Thruster-Assisted Bipedal and Legged Locomotion: Real-time QP layers coordinate ground reaction force and thrust distribution under contact constraints, enabling robust slope climbing and dynamic foot placement (Pitroda et al., 2024, Hamed et al., 2020).
• High-DOF Humanoid Control and Balancing: Large-scale QPs incorporate approximate value functions (e.g., TVLQR) for high-fidelity trajectory tracking with joint, contact, and friction constraints, achieving high-frequency control (Kuindersma et al., 2013, Hamed et al., 2020).
• Safety in Multi-Agent Systems: Coordinated avoidance, guaranteed by CBF-QPs with Lipschitz regularity, has been demonstrated in aerial and ground vehicles under actuation uncertainty and high interaction complexity (Wu et al., 2022, Frauenfelder et al., 2023).
• Cooperative Transport and Distributed Control: Strongly convex QP-based velocity and force sharing enables Lipschitz-continuous, globally convergent object motion under inter-agent force constraints (Wu et al., 14 Dec 2025).
• Learning-Based and Differentiable Layers: Neural ODEs incorporating differentiable QP layers enable end-to-end training of safe and stable controllers, supporting direct optimization and differentiable constraint learning (Miao et al., 23 Apr 2025).
• Robust Task-Space Control: Integral-augmented QP controllers provide closed-loop robustness to unmodeled dynamics in kinematically-controlled manipulators and humanoids (Djeha et al., 2023).

7. Advanced Directions and Recent Innovations

The field is advancing QP controller capabilities along several axes:

• Region-of-Attraction Enlargement: State-dependent barrier gains, synthesized via sum-of-squares (SOS) programs, can enlarge certified invariant sets and increase robustness to disturbance (Lin et al., 28 Apr 2025).
• Recursive Parameter Adaptation: Recursively feasible, bi-level adaptation strategies permit online optimization of class-$\mathcal{K}$ parameters with formal safety guarantees through differentiable QP sensitivity analysis (Parwana et al., 2021).
• Sampled-Data and Hybrid Dynamics: QCQP extensions bridge continuous-time design with discrete-time and sampled-data implementations, establishing practical stability for embedded controllers (Taylor et al., 2021).
• Task/Constraint Markovization: QPs encode multi-stage and temporal logic constraints, facilitating complex spatio-temporal task execution (Garg et al., 2019, Black et al., 2020).
• Compatibility and Dynamic Geometry: Ensuring global convergence and safety under multiple CBFs via real-time QP-controlled geometric adaptation of CLFs (Reis et al., 2024).
• Feasibility and Mode Selection in Constrained Environments: Efficient LP-based feasibility solvers dynamically reconfigure soft constraint sets to maintain controller viability (Rousseas et al., 17 Feb 2025).

---

References:

(Pitroda et al., 2024): Quadratic Programming Optimization for Bio-Inspired Thruster-Assisted Bipedal Locomotion on Inclined Slopes (Kuindersma et al., 2013): An Efficiently Solvable Quadratic Program for Stabilizing Dynamic Locomotion (Garg et al., 2019): Fixed-time Control under Spatiotemporal and Input Constraints: A Quadratic Program Based Approach (Wu et al., 2022): Quadratic Programming for Continuous Control of Safety-Critical Multi-Agent Systems Under Uncertainty (Wu et al., 14 Dec 2025): Quadratic-Programming-based Control of Multi-Robot Systems for Cooperative Object Transport (Rousseas et al., 17 Feb 2025): Feasibility Evaluation of Quadratic Programs for Constrained Control (Agrawal et al., 25 Aug 2025): Reformulations of Quadratic Programs for Lipschitz Continuity (Reis et al., 2024): On the Stability of Undesirable Equilibria in the Quadratic Program Framework for Safety-Critical Control (Miao et al., 23 Apr 2025): Opt-ODENet: A Neural ODE Framework with Differentiable QP Layers for Safe and Stable Control Design (Parwana et al., 2021): Recursive Feasibility Guided Optimal Parameter Adaptation of Differential Convex Optimization Policies for Safety-Critical Systems (Luo et al., 28 Mar 2025): Newton-PIPG: A Fast Hybrid Algorithm for Quadratic Programs in Optimal Control (Djeha et al., 2023): Robust Task-Space Quadratic Programming for Kinematic-Controlled Robots (Garg et al., 2019): Control-Lyapunov and Control-Barrier Functions based Quadratic Program for Spatio-temporal Specifications (Hamed et al., 2020): Quadrupedal Locomotion via Event-Based Predictive Control and QP-Based Virtual Constraints (Black et al., 2020): A Quadratic Program based Control Synthesis under Spatiotemporal Constraints and Non-vanishing Disturbances (Frauenfelder et al., 2023): Decentralized Vehicle Coordination and Lane Switching without Switching of Controllers (Lin et al., 28 Apr 2025): Modified Control Barrier Function for Quadratic Program Based Control Design via Sum-of-Squares Programming (Bishop et al., 2023): ReLU-QP: A GPU-Accelerated Quadratic Programming Solver for Model-Predictive Control (Taylor et al., 2021): Sampled-Data Stabilization with Control Lyapunov Functions via Quadratically Constrained Quadratic Programs