Real-Time QP Solvers: A Concise Review and Practical Guide Towards Legged Robots
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "Real-Time QP Solvers".
Title and authors: Rosa: We just discussed how we choose solvers based on whether we need sparse methods for long planning or dense methods for high frequency control, and now let’s look at what the authors actually summarize in this "Real-Time QP Solvers: A Concise Review and Practical Guide Towards Legged Robots" paper says about these solvers.
Dev: It essentially boils down to classifying the problem type first, then matching it to the right solver family. They lay out four main families: active set methods, interior-point methods, operator splitting schemes like ADMM, and augmented Lagrangian or proximal approaches.
Taro: So what’s the main point they are making about those families? Are they all equally good for every robot task?
Rosa: Not at all. The paper shows that each family has its own computational characteristics, and it's important to see how well a solver can exploit the problem structure or if it supports warm-starting.
Dev: They give us specific examples: interior-point methods are often the top choice for sparse optimal control problems in optimal control form because they exploit that sparsity really well, as seen with HPIPM fifteen.
Taro: So, if we're dealing with a long horizon problem where things become sparse, using something that exploits structure sounds like it would save us time during those critical planning steps.
Rosa: That’s right. But they also switch gears to Whole-Body Control, which usually involves smaller to medium dense problems solved at very high rates, like a hundred to a thousand Hertz six.
The paper's summary: Rosa: We’ve established the core guidance: match your solver family to your QP structure, but now we’re looking at how the authors suggest making those tools even better in their proposed improvements.
Dev: They aren't just listing solvers; they are telling us exactly how to use them by saying you should pick an interior-point solver like HPIPM for long horizon planning when sparsity is key.
Taro: And conversely, they recommend dense active-set solvers like qpOASES or Eiquadprog for those high frequency control loops where the problem is more crowded with constraints.
Rosa: The authors are pushing the idea of matching the solver family to the problem structure, which I think is really smart because it avoids wasting time on a solver that just doesn't fit your math.
Dev: They also suggest using code generation tools like FORCESPRO when you have a fixed structure and you need that highly optimized C implementation for hard real-time applications.
Taro: And they’re suggesting using ProxQP, which uses an augmented Lagrangian approach, because it seems to exhibit strong robustness in those contact-rich settings where things get messy.
Rosa: That robustness aspect is what I care about most when the robot misbehaves; you need something that handles unexpected physical interactions gracefully instead of just crashing.
Dev: They also brought up using dual active-set solvers like qpmad for dense whole body control, saying it can give a speedup because it has efficient memory access patterns and low overhead compared to some others.
Taro: It’s interesting that they are also pointing toward proximal interior-point solvers like PIQP when you're dealing with very ill-conditioned sparse QPs, which is a real challenge for embedded targets.
Rosa: So what this means for us in the field is that it’s not just about finding *a* solver; it’s about choosing the right mathematical approach to make sure your control loop stays stable and fast when the world throws curveballs.
Dev: And they emphasize that this alignment is crucial because it directly impacts your latency and your ability to meet those strict timing deadlines in the real world.
Taro: It gives us a concrete way to think about failure modes—if we pick a solver that isn't right for the problem, we know exactly where our instability is coming from before we even start deploying.
Rosa: So, this paper provides this actionable guidance on how to select solvers for legged robotics based on structure and performance metrics.
Dev: And they hint at future work, like looking into GPU-based solvers or distributed model predictive control integration, which shows the research isn't just stopping here.
The paper's improvements: Rosa: So we’re wrapping up this look at "Real-Time QP Solvers: A Concise Review and Practical Guide Towards Legged Robots" by summarizing how to pick a solver based on the robot's structure and its speed needs.
Dev: Right, it boils down to matching sparse interior-point methods for long horizon planning against dense active-set solvers for high frequency control loops.
Taro: It really changes how we think about deployment because it forces us to consider the solver family upfront instead of just picking whatever solver is popular in the community.
Rosa: And they give us some great practical advice, like using code generation tools when you have a fixed structure and hardware deployment is the main constraint.
Dev: They also showed how things like Solve Frequency per Watt can give us a better picture of energy usage, which is important when we’re trying to keep the robot running on battery power.
Taro: I think what this means for real-world testing is that we need to be very careful about the numerical stability and fault diagnosis when we push these solvers outside of a controlled lab setting.
Rosa: That’s true; it shows that even with the best solvers, the world misbehaves and you have to know how your system recovers gracefully under those real-world conditions.
Dev: So, for anyone working on real-time loops, this paper gives you a clear framework for choosing between methods when latency and failure modes are your biggest concerns.
Taro: It’s useful because it breaks down the complexity into manageable choices based on what the robot actually does—whether it's planning a long path or reacting to a sudden push.
Rosa: This paper really provides that roadmap for selecting tools—sparse structured interior points for long horizon planning and dense active-set solvers for high frequency control—to advance agile autonomous legged systems.
Dev: It’s a solid review of the landscape for real-time QP solving in robotics, and it shows that effective performance comes from aligning those solver families with the specific structure of the QP.
Taro: It sets a good baseline for us to test our own recovery strategies when we move these concepts from theory into actual field trials.
Rosa: Yeah, this is a great summary of how to get those optimization problems running reliably in real-time on embedded platforms.
Conclusion: Rosa: We’ve looked at this paper called "Real-Time QP Solvers: A Concise Review and Practical Guide Towards Legged Robots," which really shows how to pick the right mathematical tool for a robot's movement or control task based on what it needs to do and what its hardware can handle.
Dev: Yeah, it boils down to matching sparse interior-point methods for long horizon planning against dense active-set solvers for high frequency control loops.
Taro: It really changes how we think about deployment because it forces us to consider the solver family upfront instead of just picking whatever solver is popular in the community.
Rosa: And they give us some great practical advice, like using code generation tools when you have a fixed structure and hardware deployment is the main constraint.
Dev: They also showed how things like Solve Frequency per Watt can give us a better picture of energy usage, which is important when we’re trying to keep the robot running on battery power.
Taro: I think what this means for real-world testing is that we need to be very careful about the numerical stability and fault diagnosis when we push these solvers outside of a controlled lab setting.
Rosa: That’s true; it shows that even with the best solvers, the world misbehaves and you have to know how your system recovers gracefully under those real-world conditions.
Dev: So, for anyone working on real-time loops, this paper gives you a clear framework for choosing between methods when latency and failure modes are your biggest concerns.
Taro: It’s useful because it breaks down the complexity into manageable choices based on what the robot actually does—whether it's planning a long path or reacting to a sudden push.
Rosa: This paper really provides that roadmap for selecting tools—sparse structured interior points for long horizon planning and dense active-set solvers for high frequency control—to advance agile autonomous legged systems.
Dev: It’s a solid review of the landscape for real-time QP solving in robotics, and it shows that effective performance comes from aligning those solver families with the specific structure of the QP.
Taro: It sets a good baseline for us to test our own recovery strategies when we move these concepts from theory into actual field trials.
Rosa: Yeah, this is a great summary of how to get those optimization problems running reliably in real-time on embedded platforms.
Dev: It’s a lot of practical direction, but it also sets the stage for future work, like looking at GPU-based solvers or distributed model predictive control integration to push these boundaries further.
Taro: I'm looking forward to seeing how those next steps play out in actual field testing when we run these kinds of complex optimization problems outside the lab environment.
Rosa: Well, that’s what we’ve covered on this paper. It’s a solid review of the landscape for real-time QP solving in robotics. Thanks for tuning in to hear about it with us today.
Van-Nam Dinh
Institute of Science and Technology, Vinh University
cs.RO
Submitted: 2025-10-17
Updated: 2026-11-18
Comments: 12 pages, 1 figure, 2 tables
Code: https://github.com/stack-of-tasks/eiquadprog
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 79/100
The gist: Quadratic programming (QP) underpins real-time robotics by enabling efficient, constrained optimization in state estimation, motion planning…
Key concepts
- Quadratic Programming (QP)
- QP is a mathematical optimization technique used extensively in robotics to find the best solution among many possibilities. It involves minimizing or maximizing a quadratic function subject to various constraints like joint limits or physical contact rules. It is the core engine for real-time decisions in legged robots.
- Solver Classification
- Solvers are categorized into families based on their mathematical approach, such as interior-point methods (which explore the solution space from the inside) and active-set strategies (which iteratively identify active constraints). This classification helps determine which algorithm is best suited for a specific robot task, balancing speed against accuracy.
- SFPW (Solve Frequency per Watt)
- SFPW is a specialized metric that measures how efficiently a QP solver uses hardware resources. It combines solve time with power consumption to reveal the true cost of using a specific solver on an embedded system. This helps practitioners choose solvers that are not just fast, but also energy-efficient for battery-powered robots.
Terminology
Summary
Quadratic programming (QP) underpins real-time robotics by enabling efficient, constrained optimization in state estimation, motion planning, and control In legged locomotion and manipulation, essential modules like inverse dynamics, Model Predictive Control (MPC), and Whole-Body Control (WBC) are inherently QP-based, demanding reliable solutions amid tight timing, energy, and computational resources on embedded platforms. p. 0 Quadratic programming (QP) underpins real-time robotics
by enabling efficient, constrained optimization in state estimation, motion planning, and control. In legged locomotion and manipulation, essential modules like inverse dynamics, Model Predictive Control (MPC),
and Whole-Body Control (WBC) are inherently QP-based, demanding
reliable solutions amid tight timing, energy, and computational resources
on embedded platforms. Quadratic programming (QP) underpins real-time robotics
by enabling efficient, constrained optimization in state estimation, motion planning, and control. In legged locomotion and manipulation, essential modules like inverse dynamics, Model Predictive Control (MPC),
and Whole-Body Control (WBC) are inherently QP-based, demanding
reliable solutions amid tight timing, energy, and computational resources
on embedded platforms. In legged locomotion and manipulation, essential modules like inverse dynamics, Model Predictive Control (MPC),
and Whole-Body Control (WBC) are inherently QP-based, demanding
reliable solutions amid tight timing, energy, and computational resources
on embedded platforms. This paper presents a comprehensive analysis
and benchmarking study of QP solvers for legged robotics. In legged locomotion and manipulation, essential modules like inverse dynamics, Model Predictive Control (MPC),
and Whole-Body Control (WBC) are inherently QP-based, demanding
reliable solutions amid tight timing, energy, and computational resources
on embedded platforms. This paper presents a comprehensive analysis
and benchmarking study of QP solvers for legged robotics.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Solver Classification
The paper classifies solvers into principal algorithmic approaches, including interior-point methods, active-set strategies, operator-splitting schemes, and augmented Lagrangian/proximal approaches. Each solver is examined in terms of algorithmic structure, computational characteristics, and its ability to exploit problem structure and warm-starting. p. 0 We begin
by formulating the standard convex QP and classify solvers into principal algorithmic approaches: interior-point methods, active-set strategies, operator-splitting schemes, and augmented Lagrangian/proximal
approaches, while also discussing solver code generation for fixed-structure
QPs. Each solver is examined in terms of algorithmic structure, computational characteristics, and its ability to exploit problem structure and
warm-starting.Where the paper says it
Algorithmic Families
The four major algorithmic families commonly used for QPs in robotics are active-set methods, interior-point methods, operator-splitting methods (e.g., ADMM), and augmented Lagrangian/proximal methods. p. 2 There are four major algorithmic families commonly used for QPs in robotics: active-set methods, interior-point methods, operator-splitting methods (e. , ADMM), and augmented Lagrangian/proximal
methods [8]. In addition, code generation tools can produce specialized C
solvers for fixed-structure problem families, which is often attractive for embedded deployment. Active-set methods [13] iteratively identify the active inequality constraints
(CAx = uA) and solve equality-constrained subproblems:
min
x∈Rn
1
2
x
⊤Hx + q
⊤x
s. Interior-point methods (IPMs) [15] approach the feasible region from
the interior by introducing a barrier term (assuming inequalities are expressed
as Cx ≤ u):
min
x∈Rn
1
2
x
⊤Hx + q
⊤x − µ
Xni
i=1
log(ui − (Cx)i)
s. Ax = b,
(3)
where µ > 0 is the barrier parameter and strict feasibility Cx < u is required.Where the paper says it
Where the paper says it
Where the paper says it
Benchmarking Insights
Performance is reviewed using publicly available benchmarks, with a focus on metrics such as computation time, constraint satisfaction, and robustness under perturbations. The study uses Solve Frequency per Watt (SFPW) alongside standard timing and feasibility metrics to highlight solver–hardware interactions. p. 1 , ARM versus x86) [1]. In this paper, we use SFPW
alongside standard timing and feasibility metrics to highlight solver–hardware
interactions that are otherwise obscured by raw solve-time numbers. In the MPC benchmarks [1], implemented using the acados framework [12],
the authors evaluated only HPIPM, OSQP, and qpOASES. Among these, HPIPM
consistently achieved the lowest solve times for sparse QP formulations due to
its ability to exploit structured sparsity arising from the linearized system dynamics. In the WBC benchmarks [1], which involve smaller, dense QPs for whole-body
inverse dynamics, the solvers PROXQP, Eiquadprog, and qpSWIFT were evaluated
using the ARC-OPT control stack. Eiquadprog, based on a classical active-set
method, achieved sub-millisecond solve times across all tested platforms, making it well-suited for high-frequency torque control.Where the paper says it
Where the paper says it
Where the paper says it
Practical Selection Guidance
The findings emphasize the synergy between solvers, tasks, and hardware—e.g., sparse structured IPMs for long-horizon MPC and dense active-set for high-frequency WBC to advance agile, autonomous legged systems. p. 0 Our findings emphasize the synergy between solvers, tasks, and hardware—e. , sparse structured IPMs
for long-horizon MPC and dense active-set for high-frequency WBC to
advance agile, autonomous legged systems, with emerging trends toward
ill-conditioned, conic, and code-generated deployments. In summary, the study concludes that sparse solvers such as HPIPM are
well suited for large-horizon MPC tasks, where optimal-control-induced sparsity
can be exploited effectively. In contrast, dense solvers such as Eiquadprog and
PROXQP are preferable for small, high-frequency control problems like WBC. In summary, the study concludes that sparse solvers such as HPIPM are
well suited for large-horizon MPC tasks, where optimal-control-induced sparsity
can be exploited effectively. In contrast, dense solvers such as Eiquadprog and
PROXQP are preferable for small, high-frequency control problems like WBC. When the QP structure is fixed
(dimensions and sparsity pattern constant) and only parameters change online,
code-generation tools such as CVXGEN [22], CVXPYgen [23], and FORCESPRO [24]
can produce specialized C solvers with tight footprints and predictable timing. When the formulation involves conic constraints (beyond standard QP inequalities) or certificates are desired, conic IPMs such as Clarabel [25] provide a
general convex optimization route that includes QPs as a special case, albeit
with potential overhead compared to QP-specialized solvers.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
Overall, existing benchmark evidence indicates that no single solver dominates across all scenarios. Effective real-time performance emerges from aligning solver families with QP structure, accuracy targets, hardware characteristics, and energy constraints. p. 9 Overall, existing benchmark evidence indicates that no single solver dominates across all scenarios. Instead, effective real-time performance emerges from
aligning solver families with QP structure, accuracy targets, hardware characteristics, and energy constraints, underscoring the need for application-aware solver
selection in legged robotic systems. The qpmad_benchmark [19] provides a complementary and targeted evaluation of dense QP solvers commonly used in robotics, particularly for condensed
MPC and inverse-dynamics formulations. It benchmarks qpmad[10], a lightweight
C++ solver implementing the Goldfarb–Idnani dual active-set method, against
qpOASES and Eiquadprog.Where the paper says it
Where the paper says it
References
[1] Stark, F., Winkler, A., Mansard, N., et al.: Benchmarking Different QP Formulations and Solvers for Dynamic Quadrupedal Walking. p. 8 Following the study in [1], a detailed benchmarking of QP formulations
and solvers was conducted to evaluate their suitability for real-time control in
quadrupedal locomotion. The authors tested six widely used solvers—HPIPM,
PROXQP, OSQP, qpOASES, Eiquadprog, and qpSWIFT—on two canonical control
problems: Model Predictive Control (MPC) for trajectory optimization, and
Whole-Body Control (WBC) for torque-level inverse dynamics under contact
constraints. Quadratic programming (QP) [8] enables the formulation of control objectives
and constraints in a convex, computationally efficient, and theoretically grounded
framework, and it is widely adopted in legged and humanoid control stacks under
stringent latency and energy constraints [11,12]. In a standard legged robot control architecture [5], a Model Predictive Controller (MPC) computes contact forces or center-of-mass trajectories over a
arXiv:2510.Where the paper says it
Where the paper says it
[8] Bambade, A., Carpentier, J., Mansard, N., et al.: ProxQP: Yet another Quadratic Programming Solver for Robotics and beyond. p. 0 vn
Abstract. Quadratic programming (QP) underpins real-time robotics
by enabling efficient, constrained optimization in state estimation, motion planning, and control. vn
Abstract. Quadratic programming (QP) underpins real-time robotics
by enabling efficient, constrained optimization in state estimation, motion planning, and control.Where the paper says it
Where the paper says it
[9] Mastalli, C., Saurel, G., Mifsud, M., et al.: Eiquadprog: A fast and open-source implementation of the Goldfarb–Idnani solver for robotics. p. 9 The qpmad_benchmark [19] provides a complementary and targeted evaluation of dense QP solvers commonly used in robotics, particularly for condensed
MPC and inverse-dynamics formulations. It benchmarks qpmad[10], a lightweight
C++ solver implementing the Goldfarb–Idnani dual active-set method, against
qpOASES and Eiquadprog. The qpmad_benchmark [19] provides a complementary and targeted evaluation of dense QP solvers commonly used in robotics, particularly for condensed
MPC and inverse-dynamics formulations. It benchmarks qpmad[10], a lightweight
C++ solver implementing the Goldfarb–Idnani dual active-set method, against
qpOASES and Eiquadprog. Each solver is examined in terms of algorithmic structure, computational characteristics, and its ability to exploit problem structure and
warm-starting. Performance is reviewed using publicly available benchmarks, with a focus on metrics such as computation time, constraint
satisfaction, and robustness under perturbations. Each solver is examined in terms of algorithmic structure, computational characteristics, and its ability to exploit problem structure and
warm-starting. Performance is reviewed using publicly available benchmarks, with a focus on metrics such as computation time, constraint
satisfaction, and robustness under perturbations. Each solver is examined in terms of algorithmic structure, computational characteristics, and its ability to exploit problem structure and
warm-starting. Performance is reviewed using publicly available benchmarks, with a focus on metrics such as computation time, constraint
satisfaction, and robustness under perturbations. Each solver is examined in terms of algorithmic structure, computational characteristics, and its ability to exploit problem structure and
warm-starting. Performance is reviewed using publicly available benchmarks, with a focus on metrics such as computation time, constraint
satisfaction, and robustness under perturbations.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[10] Asherikov, A.: qpmad: A lightweight C++ implementation of the Goldfarb–Idnani active-set solver. p. 5 qpmad [10] Dual AS (Goldfarb–
Idnani)
✓ ✓ ✓ ✓ Dense WBC; condensed MPC
dense QPs
Lightweight, low-overhead C++ implementation;
competitive on dense repeated solves. Like other
active-set methods, can exhibit spikes if the active
set changes abruptly. qpmad [10] Dual AS (Goldfarb–
Idnani)
✓ ✓ ✓ ✓ Dense WBC; condensed MPC
dense QPs
Lightweight, low-overhead C++ implementation;
competitive on dense repeated solves. Like other
active-set methods, can exhibit spikes if the active
set changes abruptly. Each solver is examined in terms of algorithmic structure, computational characteristics, and its ability to exploit problem structure and
warm-starting. Performance is reviewed using publicly available benchmarks, with a focus on metrics such as computation time, constraint
satisfaction, and robustness under perturbations. Each solver is examined in terms of algorithmic structure, computational characteristics, and its ability to exploit problem structure and
warm-starting. Performance is reviewed using publicly available benchmarks, with a focus on metrics such as computation time, constraint
satisfaction, and robustness under perturbations.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[13] Ferreau, H.J., Kirches, C., Potschka, A., Bock, H.G., Diehl, M.: qpOASES: A parametric active-set algorithm for quadratic programming. p. 5 Backend choice
(dense/sparse) and scaling still matter; ecosystem newer than legacy AS solvers. qpOASES [13] Parametric AS ✓ ✓ ✗ ✓ Small/medium dense MPC/WBC
with stable active sets
Excellent warm-start; often very fast when active set
is stable. Backend choice
(dense/sparse) and scaling still matter; ecosystem newer than legacy AS solvers. qpOASES [13] Parametric AS ✓ ✓ ✗ ✓ Small/medium dense MPC/WBC
with stable active sets
Excellent warm-start; often very fast when active set
is stable.Where the paper says it
Where the paper says it
[14] Stellato, B., Banjac, G., Goulart, P., Bemporad, A., Boyd, S.: OSQP: An Operator Splitting Solver for Quadratic Programs. p. 1 Recent benchmarking efforts highlight that solver selection is strongly scenarioand structure-dependent [1,18,19]. For sparse MPC QPs in OCP form, structureexploiting interior-point solvers such as HPIPM [15] are often a top choice; for
large generic sparse QPs or when moderate accuracy suffices, operator-splitting
solvers such as OSQP [14] are attractive due to warm-starting and predictable
iteration costs. Recent benchmarking efforts highlight that solver selection is strongly scenarioand structure-dependent [1,18,19]. For sparse MPC QPs in OCP form, structureexploiting interior-point solvers such as HPIPM [15] are often a top choice; for
large generic sparse QPs or when moderate accuracy suffices, operator-splitting
solvers such as OSQP [14] are attractive due to warm-starting and predictable
iteration costs. vn
Abstract. Quadratic programming (QP) underpins real-time robotics
by enabling efficient, constrained optimization in state estimation, motion planning, and control. vn
Abstract. Quadratic programming (QP) underpins real-time robotics
by enabling efficient, constrained optimization in state estimation, motion planning, and control.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[15] Frison, G., Diehl, M.: HPIPM: A high-performance quadratic programming framework for MPC. p. 0 Quadratic programming (QP) [8] enables the formulation of control objectives
and constraints in a convex, computationally efficient, and theoretically grounded
framework, and it is widely adopted in legged and humanoid control stacks under
stringent latency and energy constraints [11,12]. In a standard legged robot control architecture [5], a Model Predictive Controller (MPC) computes contact forces or center-of-mass trajectories over a
arXiv:2510. Quadratic programming (QP) [8] enables the formulation of control objectives
and constraints in a convex, computationally efficient, and theoretically grounded
framework, and it is widely adopted in legged and humanoid control stacks under
stringent latency and energy constraints [11,12]. In a standard legged robot control architecture [5], a Model Predictive Controller (MPC) computes contact forces or center-of-mass trajectories over a
arXiv:2510.Where the paper says it
Where the paper says it
[18] Caron, S., Zaki, A., Otta, P., et al.: qpbenchmark: Benchmark for quadratic programming solvers available in Python. p. 1 Second, we consolidate publicly available benchmarking evidence using tools such as
qpbenchmark [18] and qpmad_benchmark [19], and we translate these results into
practical solver-selection guidance for legged robot MPC and WBC pipelines. 2 Quadratic Programming Optimization Problem
We use n for the number of decision variables, ne for the number of equality
constraints, and ni for the number of inequality constraints. Second, we consolidate publicly available benchmarking evidence using tools such as
qpbenchmark [18] and qpmad_benchmark [19], and we translate these results into
practical solver-selection guidance for legged robot MPC and WBC pipelines. 2 Quadratic Programming Optimization Problem
We use n for the number of decision variables, ne for the number of equality
constraints, and ni for the number of inequality constraints.Where the paper says it
Where the paper says it
[19] Asherikov, A.: qpmad benchmark: Benchmarking dense QP solvers for robotics. p. 1 First,
we review QP solver families relevant to real-time control in robotics and summarize representative solvers across open-source and commercial ecosystems. Second, we consolidate publicly available benchmarking evidence using tools such as
qpbenchmark [18] and qpmad_benchmark [19], and we translate these results into
practical solver-selection guidance for legged robot MPC and WBC pipelines. First,
we review QP solver families relevant to real-time control in robotics and summarize representative solvers across open-source and commercial ecosystems. Second, we consolidate publicly available benchmarking evidence using tools such as
qpbenchmark [18] and qpmad_benchmark [19], and we translate these results into
practical solver-selection guidance for legged robot MPC and WBC pipelines.Where the paper says it
Where the paper says it
[20] Arnström, D., Bemporad, A., Axehill, D.: A Dual Active-Set Solver for Embedded Quadratic Programming Using Recursive LDL⊤ Updates. p. 9 DAQP [20] is most relevant when MPC QPs are fully condensed into
dense problems with fixed structure across time steps, enabling aggressive warmstarting and efficient recursive LDL⊤ updates; it is not intended for large, generic
sparse QPs. PIQP [21] targets sparse (and dense) convex QPs with an emphasis on ill-conditioning, combining a proximal mechanism with an interior-point
core, and is therefore attractive when numerical stability, regularization, and
allocation-free updates are critical on embedded targets. Quadratic programming (QP) [8] enables the formulation of control objectives
and constraints in a convex, computationally efficient, and theoretically grounded
framework, and it is widely adopted in legged and humanoid control stacks under
stringent latency and energy constraints [11,12]. In a standard legged robot control architecture [5], a Model Predictive Controller (MPC) computes contact forces or center-of-mass trajectories over a
arXiv:2510.Where the paper says it
Where the paper says it
[21] Schwan, R., Jiang, Y., Kuhn, D., Jones, C.N.: PIQP: A Proximal Interior-Point Quadratic Programming Solver. p. 1 For dense WBC QPs, active-set solvers such as qpOASES [13],
Eiquadprog [9], and the lightweight qpmad [10] remain strong baselines. Beyond these widely used tools, modern solver options include embedded-focused
dual active-set solvers such as DAQP [20], proximal interior-point solvers designed
for ill-conditioned sparse QPs such as PIQP [21], and sparse primal-dual IPMs
positioned for robotics such as qpSWIFT [16]. Quadratic programming (QP) [8] enables the formulation of control objectives
and constraints in a convex, computationally efficient, and theoretically grounded
framework, and it is widely adopted in legged and humanoid control stacks under
stringent latency and energy constraints [11,12]. In a standard legged robot control architecture [5], a Model Predictive Controller (MPC) computes contact forces or center-of-mass trajectories over a
arXiv:2510.Where the paper says it
Where the paper says it
[25] Goulart, P.J., Chen, Y.: Clarabel: An Interior-Point Solver for Conic Programs with Quadratic Objectives. p. 6 Best when dimensions/structure are fixed and modest; less flexible if
formulation changes frequently. Clarabel [25] IP (conic; QP-capable) ✓ ✓ ✓ ✓ When conic constraints/certificates
are needed
General convex conic solver with quadratic objectives; homogeneous embedding approach. Best when dimensions/structure are fixed and modest; less flexible if
formulation changes frequently. Clarabel [25] IP (conic; QP-capable) ✓ ✓ ✓ ✓ When conic constraints/certificates
are needed
General convex conic solver with quadratic objectives; homogeneous embedding approach. Quadratic programming (QP) [8] enables the formulation of control objectives
and constraints in a convex, computationally efficient, and theoretically grounded
framework, and it is widely adopted in legged and humanoid control stacks under
stringent latency and energy constraints [11,12]. In a standard legged robot control architecture [5], a Model Predictive Controller (MPC) computes contact forces or center-of-mass trajectories over a
arXiv:2510. Quadratic programming (QP) [8] enables the formulation of control objectives
and constraints in a convex, computationally efficient, and theoretically grounded
framework, and it is widely adopted in legged and humanoid control stacks under
stringent latency and energy constraints [11,12]. In a standard legged robot control architecture [5], a Model Predictive Controller (MPC) computes contact forces or center-of-mass trajectories over a
arXiv:2510.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[3] Van Nam, D., Nguyen, M.T., Kim, G.-W., et al.: Fusion Consistency for Industrial Robot Navigation: An Integrated SLAM Framework with Multiple 2D LiDARVisual-Inertial Sensors. p. 0 Keywords: Quadratic Programming
· Optimization
· Real-Time Control · Legged Robots · QP Solvers
1 Introduction
Real-time optimization forms the backbone of advanced legged robot control [2],
including state estimation [3,4], motion planning [2], and dynamic control [5,6]. Quadratic programming (QP) [8] enables the formulation of control objectives
and constraints in a convex, computationally efficient, and theoretically grounded
framework, and it is widely adopted in legged and humanoid control stacks under
stringent latency and energy constraints [11,12]. Keywords: Quadratic Programming
· Optimization
· Real-Time Control · Legged Robots · QP Solvers
1 Introduction
Real-time optimization forms the backbone of advanced legged robot control [2],
including state estimation [3,4], motion planning [2], and dynamic control [5,6]. Quadratic programming (QP) [8] enables the formulation of control objectives
and constraints in a convex, computationally efficient, and theoretically grounded
framework, and it is widely adopted in legged and humanoid control stacks under
stringent latency and energy constraints [11,12].Where the paper says it
Where the paper says it
[4] D. Van Nam and K. p. 0 Real-Time QP Solvers: A Concise Review and
Practical Guide Towards Legged Robots
Van-Nam Dinh
1
Institute of Science and Technology, Vinh University, Vinh, Nghe An, Viet Nam
namdv@vinhuni. vn
Abstract. Real-Time QP Solvers: A Concise Review and
Practical Guide Towards Legged Robots
Van-Nam Dinh
1
Institute of Science and Technology, Vinh University, Vinh, Nghe An, Viet Nam
namdv@vinhuni. vn
Abstract. Quadratic programming (QP) underpins real-time robotics
by enabling efficient, constrained optimization in state estimation, motion planning, and control. In legged locomotion and manipulation, essential modules like inverse dynamics, Model Predictive Control (MPC),
and Whole-Body Control (WBC) are inherently QP-based, demanding
reliable solutions amid tight timing, energy, and computational resources
on embedded platforms. Quadratic programming (QP) underpins real-time robotics
by enabling efficient, constrained optimization in state estimation, motion planning, and control. In legged locomotion and manipulation, essential modules like inverse dynamics, Model Predictive Control (MPC),
and Whole-Body Control (WBC) are inherently QP-based, demanding
reliable solutions amid tight timing, energy, and computational resources
on embedded platforms. Real-Time QP Solvers: A Concise Review and
Practical Guide Towards Legged Robots
Van-Nam Dinh
1
Institute of Science and Technology, Vinh University, Vinh, Nghe An, Viet Nam
namdv@vinhuni. vn
Abstract. Real-Time QP Solvers: A Concise Review and
Practical Guide Towards Legged Robots
Van-Nam Dinh
1
Institute of Science and Technology, Vinh University, Vinh, Nghe An, Viet Nam
namdv@vinhuni. vn
Abstract.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[5] Di Carlo, J., Wensing, P.M., Katz, B., Bledt, G., Kim, S.: Dynamic Locomotion in the MIT Cheetah 3 Through Convex Model-Predictive Control. p. 0 Quadratic programming (QP) underpins real-time robotics
by enabling efficient, constrained optimization in state estimation, motion planning, and control. In legged locomotion and manipulation, essential modules like inverse dynamics, Model Predictive Control (MPC),
and Whole-Body Control (WBC) are inherently QP-based, demanding
reliable solutions amid tight timing, energy, and computational resources
on embedded platforms. Quadratic programming (QP) underpins real-time robotics
by enabling efficient, constrained optimization in state estimation, motion planning, and control. In legged locomotion and manipulation, essential modules like inverse dynamics, Model Predictive Control (MPC),
and Whole-Body Control (WBC) are inherently QP-based, demanding
reliable solutions amid tight timing, energy, and computational resources
on embedded platforms. , sparse structured IPMs
for long-horizon MPC and dense active-set for high-frequency WBC to
advance agile, autonomous legged systems, with emerging trends toward
ill-conditioned, conic, and code-generated deployments. Keywords: Quadratic Programming
· Optimization
· Real-Time Control · Legged Robots · QP Solvers
1 Introduction
Real-time optimization forms the backbone of advanced legged robot control [2],
including state estimation [3,4], motion planning [2], and dynamic control [5,6]. , sparse structured IPMs
for long-horizon MPC and dense active-set for high-frequency WBC to
advance agile, autonomous legged systems, with emerging trends toward
ill-conditioned, conic, and code-generated deployments. Keywords: Quadratic Programming
· Optimization
· Real-Time Control · Legged Robots · QP Solvers
1 Introduction
Real-time optimization forms the backbone of advanced legged robot control [2],
including state estimation [3,4], motion planning [2], and dynamic control [5,6].Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[6] Wensing, P.M., Posa, M., Escande, A., Mansard, N., Del Prete, A.: OptimizationBased Control for Dynamic Legged Robots. p. 0 , sparse structured IPMs
for long-horizon MPC and dense active-set for high-frequency WBC to
advance agile, autonomous legged systems, with emerging trends toward
ill-conditioned, conic, and code-generated deployments. Keywords: Quadratic Programming
· Optimization
· Real-Time Control · Legged Robots · QP Solvers
1 Introduction
Real-time optimization forms the backbone of advanced legged robot control [2],
including state estimation [3,4], motion planning [2], and dynamic control [5,6]. , sparse structured IPMs
for long-horizon MPC and dense active-set for high-frequency WBC to
advance agile, autonomous legged systems, with emerging trends toward
ill-conditioned, conic, and code-generated deployments. Keywords: Quadratic Programming
· Optimization
· Real-Time Control · Legged Robots · QP Solvers
1 Introduction
Real-time optimization forms the backbone of advanced legged robot control [2],
including state estimation [3,4], motion planning [2], and dynamic control [5,6].Where the paper says it
Where the paper says it
[7] Farshidian, F., et al.: OCS2: An Open Source Library for Optimal Control of Switched Systems. p. 1 Building on this context, this paper makes two primary contributions. First,
we review QP solver families relevant to real-time control in robotics and summarize representative solvers across open-source and commercial ecosystems. Building on this context, this paper makes two primary contributions. First,
we review QP solver families relevant to real-time control in robotics and summarize representative solvers across open-source and commercial ecosystems.Where the paper says it
Where the paper says it
[11] Verschueren, R.: Convex Approximation Methods for Nonlinear Model Predictive Control. p. 0 Quadratic programming (QP) [8] enables the formulation of control objectives
and constraints in a convex, computationally efficient, and theoretically grounded
framework, and it is widely adopted in legged and humanoid control stacks under
stringent latency and energy constraints [11,12]. In a standard legged robot control architecture [5], a Model Predictive Controller (MPC) computes contact forces or center-of-mass trajectories over a
arXiv:2510. Quadratic programming (QP) [8] enables the formulation of control objectives
and constraints in a convex, computationally efficient, and theoretically grounded
framework, and it is widely adopted in legged and humanoid control stacks under
stringent latency and energy constraints [11,12]. In a standard legged robot control architecture [5], a Model Predictive Controller (MPC) computes contact forces or center-of-mass trajectories over a
arXiv:2510.Where the paper says it
Where the paper says it
[3] Van Nam, D., Nguyen, M.T., Kim, G.-W., et al.: Fusion Consistency for Industrial Robot Navigation: An Integrated SLAM Framework with Multiple 2D LiDARVisual-Inertial Sensors. Computers and Electrical Engineering 120, 109607 (2024)
[4] D. Van Nam and K. Gon-Woo,: Learning Observation Model for Factor Graph Based State Estimation Using Intrinsic Sensors. IEEE Transactions on Automation Science and Engineering, vol. 20, no. 3, pp. 2049-2062, July 2023, doi: 10.1109/TASE.2023
[5] Di Carlo, J., Wensing, P.M., Katz, B., Bledt, G., Kim, S.: Dynamic Locomotion in the MIT Cheetah 3 Through Convex Model-Predictive Control. In: Proc. IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS), pp. 1–9. IEEE (2018)
[6] Wensing, P.M., Posa, M., Escande, A., Mansard, N., Del Prete, A.: OptimizationBased Control for Dynamic Legged Robots. IEEE Transactions on Robotics 40, 43–63 (2023)
[7] Farshidian, F., et al.: OCS2: An Open Source Library for Optimal Control of Switched Systems. Software available at: https://github.com/leggedrobotics/ocs2 (accessed: 2025-12-12)
[8] Bambade, A., Carpentier, J., Mansard, N., et al.: ProxQP: Yet another Quadratic Programming Solver for Robotics and beyond. In: Proc. Robotics: Science and Systems (RSS) (2022)
[9] Mastalli, C., Saurel, G., Mifsud, M., et al.: Eiquadprog: A fast and open-source implementation of the Goldfarb–Idnani solver for robotics. Software available at: https://github.com/stack-of-tasks/eiquadprog (accessed: 2025-12-12)
[10] Asherikov, A.: qpmad: A lightweight C++ implementation of the Goldfarb–Idnani active-set solver. Available at: https://github.com/asherikov/qpmad (accessed: 2025-12-12)
[13] Ferreau, H.J., Kirches, C., Potschka, A., Bock, H.G., Diehl, M.: qpOASES: A parametric active-set algorithm for quadratic programming. Mathematical Programming Computation 6(4), 327–363 (2014)
[14] Stellato, B., Banjac, G., Goulart, P., Bemporad, A., Boyd, S.: OSQP: An Operator Splitting Solver for Quadratic Programs. IEEE Trans. Control Syst. Technol. 27(6), 2476–2488 (2019)
[15] Frison, G., Diehl, M.: HPIPM: A high-performance quadratic programming framework for MPC. IFAC-PapersOnLine 53(2), 6566–6571 (2020)
[17] O’Donoghue, B., Chu, E., Parikh, N., Boyd, S.: SCS: Splitting Conic Solver. p. 7 Solver (Method) Matrix Warm-S Hot-S Early Term. Dual-Gap I-H
qpOASES (AS)[13] Dense / Sparse ✓ ✓ ✗ ✓ ✗
Quadprog (AS)[26] Dense ✗ ✗ ✗ ✓ ✗
Eiquadprog (AS)[9] Dense ✓ ✗ ✗ ✓ ✗
qpmad (AS)[10] Both ✓ ✓ ✓ ✓ ✗
DAQP (AS)[20] Both ✓ ✓ ✓ ✓ ✗
Gurobi (IP)[27] Sparse ✗ ✗ ✓ ✓ ✗
MOSEK (IP)[28] Sparse ✗ ✗ ✓ ✓ ✗
HPIPM (IP)[15] Sparse (OCP form) ✓ ✓ ✓ ✓ ✗
OSQP (OS)[14] Sparse ✓ ✓ ✓ ✗ (no explicit gap) ✗
SCS (OS)[17] Sparse ✓ ✓ ✓ ✗ ✗
ProxQP (AL)[8] Both ✓ ✓ ✓ ✓ ✓
The qpbenchmark suite [18] includes convex QPs derived from linear MPC
tasks such as cart-pole balancing and humanoid walking. vn
Abstract. Quadratic programming (QP) underpins real-time robotics
by enabling efficient, constrained optimization in state estimation, motion planning, and control.Where the paper says it
Where the paper says it
[18] Caron, S., Zaki, A., Otta, P., et al.: qpbenchmark: Benchmark for quadratic programming solvers available in Python. https://github.com/qpsolvers/qpbenchmark (accessed: 2025-12-12)
[19] Asherikov, A.: qpmad benchmark: Benchmarking dense QP solvers for robotics. https://github.com/asherikov/qpmad benchmark (accessed: 2025-12-12)
[24] Embotech AG: FORCESPRO Documentation (Solver Options / QP FAST, PDIP, ADMM). Available at: https://forces. p. 0 Quadratic programming (QP) [8] enables the formulation of control objectives
and constraints in a convex, computationally efficient, and theoretically grounded
framework, and it is widely adopted in legged and humanoid control stacks under
stringent latency and energy constraints [11,12]. In a standard legged robot control architecture [5], a Model Predictive Controller (MPC) computes contact forces or center-of-mass trajectories over a
arXiv:2510. For dense WBC QPs, active-set solvers such as qpOASES [13],
Eiquadprog [9], and the lightweight qpmad [10] remain strong baselines. Beyond these widely used tools, modern solver options include embedded-focused
dual active-set solvers such as DAQP [20], proximal interior-point solvers designed
for ill-conditioned sparse QPs such as PIQP [21], and sparse primal-dual IPMs
positioned for robotics such as qpSWIFT [16].Where the paper says it
Where the paper says it
[25] Goulart, P.J., Chen, Y.: Clarabel: An Interior-Point Solver for Conic Programs with Quadratic Objectives. arXiv preprint arXiv:2405.12762 (2024)
[11] Verschueren, R.: Convex Approximation Methods for Nonlinear Model Predictive Control. PhD thesis, Albert-Ludwigs-Universität Freiburg im Breisgau (2018)
[12] Verschueren, R., Frison, G., Kouzoupis, D., et al.: acados—a modular open-source framework for fast embedded optimal control. p. 5 FORCESPRO [24] Codegen + IPM / fast
QP modes
✓ ✓ ✓ ✓ Hard real-time embedded MPC
with tooling
Mature code generation and deployment workflow;
solver modes for embedded targets. Commercial licensing; reduced transparency vs open-source.Where the paper says it
[26] Goldfarb, D., Idnani, A. p. 5 Worst-case iteration variability
⇒ potential
timing jitter under contact switches/disturbances. qpmad [10] Dual AS (Goldfarb–
Idnani)
✓ ✓ ✓ ✓ Dense WBC; condensed MPC
dense QPs
Lightweight, low-overhead C++ implementation;
competitive on dense repeated solves. Solver (Method) Matrix Warm-S Hot-S Early Term. Dual-Gap I-H
qpOASES (AS)[13] Dense / Sparse ✓ ✓ ✗ ✓ ✗
Quadprog (AS)[26] Dense ✗ ✗ ✗ ✓ ✗
Eiquadprog (AS)[9] Dense ✓ ✗ ✗ ✓ ✗
qpmad (AS)[10] Both ✓ ✓ ✓ ✓ ✗
DAQP (AS)[20] Both ✓ ✓ ✓ ✓ ✗
Gurobi (IP)[27] Sparse ✗ ✗ ✓ ✓ ✗
MOSEK (IP)[28] Sparse ✗ ✗ ✓ ✓ ✗
HPIPM (IP)[15] Sparse (OCP form) ✓ ✓ ✓ ✓ ✗
OSQP (OS)[14] Sparse ✓ ✓ ✓ ✗ (no explicit gap) ✗
SCS (OS)[17] Sparse ✓ ✓ ✓ ✗ ✗
ProxQP (AL)[8] Both ✓ ✓ ✓ ✓ ✓
The qpbenchmark suite [18] includes convex QPs derived from linear MPC
tasks such as cart-pole balancing and humanoid walking.Where the paper says it
Where the paper says it
[27] Gurobi Optimization, LLC.: Gurobi Optimizer Reference Manual, 2024. p. 7 Solver (Method) Matrix Warm-S Hot-S Early Term. Dual-Gap I-H
qpOASES (AS)[13] Dense / Sparse ✓ ✓ ✗ ✓ ✗
Quadprog (AS)[26] Dense ✗ ✗ ✗ ✓ ✗
Eiquadprog (AS)[9] Dense ✓ ✗ ✗ ✓ ✗
qpmad (AS)[10] Both ✓ ✓ ✓ ✓ ✗
DAQP (AS)[20] Both ✓ ✓ ✓ ✓ ✗
Gurobi (IP)[27] Sparse ✗ ✗ ✓ ✓ ✗
MOSEK (IP)[28] Sparse ✗ ✗ ✓ ✓ ✗
HPIPM (IP)[15] Sparse (OCP form) ✓ ✓ ✓ ✓ ✗
OSQP (OS)[14] Sparse ✓ ✓ ✓ ✗ (no explicit gap) ✗
SCS (OS)[17] Sparse ✓ ✓ ✓ ✗ ✗
ProxQP (AL)[8] Both ✓ ✓ ✓ ✓ ✓
The qpbenchmark suite [18] includes convex QPs derived from linear MPC
tasks such as cart-pole balancing and humanoid walking. Each solver is examined in terms of algorithmic structure, computational characteristics, and its ability to exploit problem structure and
warm-starting. Performance is reviewed using publicly available benchmarks, with a focus on metrics such as computation time, constraint
satisfaction, and robustness under perturbations. Each solver is examined in terms of algorithmic structure, computational characteristics, and its ability to exploit problem structure and
warm-starting. Performance is reviewed using publicly available benchmarks, with a focus on metrics such as computation time, constraint
satisfaction, and robustness under perturbations.Where the paper says it
Where the paper says it
Where the paper says it
[28] MOSEK ApS.: The MOSEK Optimization Toolbox for Python Manual, Version 10. p. 7 Solver (Method) Matrix Warm-S Hot-S Early Term. Dual-Gap I-H
qpOASES (AS)[13] Dense / Sparse ✓ ✓ ✗ ✓ ✗
Quadprog (AS)[26] Dense ✗ ✗ ✗ ✓ ✗
Eiquadprog (AS)[9] Dense ✓ ✗ ✗ ✓ ✗
qpmad (AS)[10] Both ✓ ✓ ✓ ✓ ✗
DAQP (AS)[20] Both ✓ ✓ ✓ ✓ ✗
Gurobi (IP)[27] Sparse ✗ ✗ ✓ ✓ ✗
MOSEK (IP)[28] Sparse ✗ ✗ ✓ ✓ ✗
HPIPM (IP)[15] Sparse (OCP form) ✓ ✓ ✓ ✓ ✗
OSQP (OS)[14] Sparse ✓ ✓ ✓ ✗ (no explicit gap) ✗
SCS (OS)[17] Sparse ✓ ✓ ✓ ✗ ✗
ProxQP (AL)[8] Both ✓ ✓ ✓ ✓ ✓
The qpbenchmark suite [18] includes convex QPs derived from linear MPC
tasks such as cart-pole balancing and humanoid walking.Where the paper says it
[29] Tuna, Turcan, et al.: Informed, constrained, aligned: A field analysis on degeneracyaware point cloud registration in the wild. p. 0 This paper presents a comprehensive analysis
and benchmarking study of QP solvers for legged robotics. We begin
by formulating the standard convex QP and classify solvers into principal algorithmic approaches: interior-point methods, active-set strategies, operator-splitting schemes, and augmented Lagrangian/proximal
approaches, while also discussing solver code generation for fixed-structure
QPs. 21773v2 [cs. RO] 12 Dec 2025Where the paper says it
Where the paper says it
--- Page 1 ---
Real-Time QP Solvers: A Concise Review and Practical Guide Towards Legged Robots
Abstract. Quadratic programming (QP) underpins real-time robotics by enabling efficient, constrained optimization in state estimation, motion planning, and control. p. 0 vn
Abstract. Quadratic programming (QP) underpins real-time robotics
by enabling efficient, constrained optimization in state estimation, motion planning, and control.Where the paper says it
--- Page 2 ---
finite horizon, often yielding sparse optimal-control-structured QPs [7,12] In contrast, a Whole-Body Controller (WBC) converts these high-level commands into joint torques or velocities subject to actuation, friction, and contact constraints, typically resulting in small-to-medium dense QPs solved at high rates (100–1000 Hz) [6] In both cases, solver speed, memory resources, numerical robustness, and warm-starting determine closed-loop feasibility
--- Page 3 ---
In both cases, solver speed, memory resources, numerical robustness, and warm-starting determine closed-loop feasibility
--- Page 4 ---
For sparse MPC QPs in OCP form, structure exploiting interior-point solvers such as HPIPM [15] are often a top choice; for large generic sparse QPs or when moderate accuracy suffices, operator-splitting solvers such as OSQP [14] are attractive due to warm-starting and predictable iteration costs For dense WBC QPs, active-set solvers such as qpOASES [13], Eiquadprog [9], and the lightweight qpmad [10] remain strong baselines Beyond these widely used tools, modern solver options include embedded-focused dual active-set solvers such as DAQP [20], proximal interior-point solvers designed for ill-conditioned sparse QPs such as PIQP [21], and sparse primal-dual IPMs positioned for robotics such as qpSWIFT [16] In industrial deployment, codegenerated solvers (e.g., FORCESPRO [24], CVXGEN [22], CVXPYgen [23]) can provide highly optimized C implementations for fixed-structure QPs Finally, when conic constraints or certificates are required, conic IPMs such as Clarabel [25] provide a general convex optimization route that also covers QPs
--- Page 5 ---
Table 1: Unified overview of QP solver families and representative solvers for real-time robotics. AS = active set; AL = Augmented Lagrangian; IP = Interior-Point; OS = Operator Splitting; WS = warm-start; Acc = high-accuracy capability (when configured accordingly); Early = early termination / inexact solve support; RT = practical real-time suitability (scenario dependent). p. 4 Table 1: Unified overview of QP solver families and representative solvers for real-time robotics. AS = active set; AL =
Augmented Lagrangian; IP = Interior-Point; OS = Operator Splitting; WS = warm-start; Acc = high-accuracy capability
(when configured accordingly); Early = early termination / inexact solve support; RT = practical real-time suitability (scenariodependent).Where the paper says it
--- Page 6 ---
Family / Solver Method family WS Acc Early RT Best scenarios Key notes (advantages / limitations)
(A) Algorithm-family summary
Direct (eq.-only) KKT / linear solve N/A Equality-only QPs; fixed structure Extremely fast when applicable; limited because robotics typically requires inequalities (contacts, friction, torque/state limits). p. 4 Family / Solver Method family WS Acc Early RT Best scenarios Key notes (advantages / limitations)
(A) Algorithm-family summary
Direct (eq. -only) KKT / linear solve N/A ✓ ✗ ✓ Equality-only QPs; fixed structure Extremely fast when applicable; limited because
robotics typically requires inequalities (contacts, friction, torque/state limits).Where the paper says it
(B) Representative solvers (selection-oriented)
HPIPM [15] IP, structureexploiting Sparse OCP/MPC QPs (blockbanded/tree) Real-time feasible when OCP sparsity is exploited via customized linear algebra; less ideal if sparsity does not match (or if condensing destroys structure).
--- Page 7 ---
Family / Solver Method family WS Acc Early RT Best scenarios Key notes (advantages / limitations)
--- Page 8 ---
Fig. 1: Performance of selected QP solvers on a convex MPC test set. p. 7 1: Performance of selected QP solvers on a convex MPC test set. A “✗”
indicates that the solver failed to reliably solve some instances at the required
tolerance.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Implement scenario-specific solver selection based on problem structure to match solver families with QP characteristics, such as using sparse structured IPMs like HPIPM for long-horizon MPC QPs or dense active-set solvers like qpOASES for high-frequency Whole Body Control (WBC) QPs.
-
Incorporate energy normalized metrics like Solve Frequency per Watt (SFPW) alongside standard timing and feasibility metrics to highlight
solver–hardware interactions that are otherwise obscured by raw solve-time numbers,
guiding the selection of solver–hardware combinations based on energy constraints. -
Utilize code generation tools such as FORCESPRO, CVXGEN, or CVXPYgen for fixed-structure problems to produce specialized C solvers with tight footprints and predictable timing for hard real-time embedded MPC applications.
-
Deploy ProxQP when strong robustness and stable behavior across diverse robotics QPs are critical, especially in contact-rich settings, leveraging its Augmented Lagrangian approach which
exhibits strong robustness in contact-rich settings.
-
Use Dual Active-Set solvers like qpmad for dense WBC or condensed MPC formulations to achieve up to a twofold speedup over Eiquadprog on small-to-medium problems due to
efficient memory access patterns, low overhead, and minimal dependencies.
-
Integrate PIQP when numerical stability, regularization, and allocation-free updates are required for ill-conditioned sparse QPs on embedded targets.
Abstract
Quadratic programming (QP) underpins real-time robotics by enabling efficient, constrained optimization in state estimation, motion planning, and control. In legged locomotion and manipulation, essential modules like inverse dynamics, Model Predictive Control (MPC), and Whole-Body Control (WBC) are inherently QP-based, demanding reliable solutions amid tight timing, energy, and computational resources on embedded platforms. This paper presents a comprehensive analysis and benchmarking study of QP solvers for legged robotics. We begin by formulating the standard convex QP and classify solvers into principal algorithmic approaches: interior-point methods, active-set strategies, operator-splitting schemes, and augmented Lagrangian/proximal approaches, while also discussing solver code generation for fixed-structure QPs. Each solver is examined in terms of algorithmic structure, computational characteristics, and its ability to exploit problem structure and warm-starting. Performance is reviewed using publicly available benchmarks, with a focus on metrics such as computation time, constraint satisfaction, and robustness under perturbations. Unified comparison tables yield practical guidance for solver selection, underscoring trade-offs in speed, accuracy, and energy efficiency. Our findings emphasize the synergy between solvers, tasks, and hardware -- e.g., sparse structured IPMs for long-horizon MPC and dense active-set for high-frequency WBC to advance agile, autonomous legged systems, with emerging trends toward ill-conditioned, conic, and code-generated deployments.
Sources
- Benchmarking Different QP Formulations and Solvers for Dynamic Quadrupedal Walking
- Humanoid Locomotion and Manipulation: Current Progress and Challenges in Control, Planning, and Learning
- A Dual Active-Set Solver for Embedded Quadratic Programming Using Recursive LDL' Updates
- PIQP: A Proximal Interior-Point Quadratic Programming Solver
- Clarabel: An interior-point solver for conic programs with quadratic objectives
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation