ElastiQP

Warning

Currently under construction, apologies for any rough edges

An always-feasible QP solver for constrained robot control.

ElastiQP solves the following problem:

$$ \begin{align*} \underset{x, t}{\text{minimize}} & \quad \frac{1}{2}x^TQx + q^Tx + w^T t \\ \text{s.t.} & \quad Ax = b \\ & \quad Gx - t \leq h \\ & \quad t \geq 0 \end{align*} $$

i.e, a QP with hard equality constraints $Ax = b$ and elastic inequality constraints $Gx \leq h$, where every inequality constraint is relaxed with an $\ell_1$ penalty term defined by $w > 0$.

Notably, the elastic slacks $t$ are eliminated analytically, so the condensed system stays $n \times n$ regardless of the number of constraints. This enables fast compute times, even with large numbers of inequality constraints $p$.

Why ElastiQP?

For robot control, we are typically interested in solving small-scale, dense QPs, where the number of inequality constraints $p$ may be much larger than the number of decision variables $n$. In this setting, inequality constraints can often be momentarily infeasible, but in a control loop, we always need to return a reasonable solution.

Likewise, in the control setting, we often encode dynamics via strict equality constraints. For typical systems, dynamics constraints are feasible by construction, and relaxing these would lead to unrealistic behavior.

Given this, in the case of infeasibility, ElastiQP naturally relaxes the inequality constraints in an $\ell_1$ manner, relaxing only the constraints that strictly need to be adjusted for a feasible solution.

Overview

ElastiQP is a C++/Eigen header-only library with Python bindings and a JAX foreign function interface (FFI).

ElastiQP's primary backend (das.hpp) is is a dual active-set method, based on DAQP. We also have two additional backends, which were primarily used as a point of comparison for the paper: pdal.hpp is a primal-dual augmented Lagrangian method, based on ProxQP, and ipm.hpp is a proximal interior point method, based on PIQP and inspired by qpax

ElastiQP is fast, particularly when warm-started. For a rough sense of numbers, on a laptop with an Intel i7 CPU, ElastiQP can solve humanoid-scale whole-body control problems at approximately 50 us.

Installation

C++

From source:

git clone https://github.com/StanfordASL/elastiqp
cd elastiqp
cmake -B build . -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release -j
# Optional: cmake --install build --config Release

For best performance (on your own device), you can also build with -march=native

cmake -B build-native . -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_FLAGS="-march=native"
cmake --build build-native --config Release -j

If you've installed with CMake, you can also find_package(elastiqp)

Python

From PyPI

pip install elastiqp

From source

If installing the Python bindings from source, first make sure that you've run the build described above in the C++ section. You'll need to have nanobind (and jax, if you want to use the FFI) installed in your current python venv when building. Then, from the top-level of the repo,

pip install .

JAX/PyTorch dependencies can be installed with pip install "elastiqp[jax]" or pip install "elastiqp[torch]", respectively.

Note: if using UV, you can directly replace the above pip commands with uv pip

Usage

C++

#include "elastiqp/elastiqp.hpp"
// or one backend only: "elastiqp/das.hpp", "pdal.hpp", "ipm.hpp"

// If you just need to solve a single problem
elastiqp::Solution sol = elastiqp::Solve(Q, q, A, b, G, h, penalty);
// or without equalities: elastiqp::Solve(Q, q, G, h, penalty)

// If you are solving multiple times in a control loop:
elastiqp::Solver solver;
solver.setup(Q, q, A, b, G, h, penalty);
while (running) {
  // Set your updated problem data and warm-start
  solver.set_q(q_k); solver.set_h(h_k); solver.set_b(b_k);
  const elastiqp::Solution& sol = solver.solve();
}

Python

import elastiqp

# If you just need to solve a single problem:
sol = elastiqp.solve(Q, q, G, h, penalty, A=A, b=b)
# Specify the backend (default "das") via the method kwarg
sol = elastiqp.solve(Q, q, G, h, penalty, A=A, b=b, method="das")

# If you are solving multiple times in a control loop:
solver = elastiqp.Solver()
solver.setup(Q, q, G, h, penalty, A=A, b=b)
sol = solver.solve()
solver.update(q=q_k, h=h_k, b=b_k)
sol = solver.solve()

JAX / PyTorch

Warning

Differentiability support is still in beta, and is not implemented in the default (active set) backend

import jax
jax.config.update("jax_enable_x64", True)
import elastiqp.jax

# Same API as Python, just with elastiqp.jax
sol = elastiqp.jax.solve(Q, q, G, h, penalty, A=A, b=b)

# Compatible with jax.grad and vjp on the pdal / ipm backends
def loss(q_):
    sol = elastiqp.jax.solve(Q, q_, G, h, penalty, A=A, b=b, method="ipm")
    return jnp.sum(sol.x**2)

grad_q = jax.grad(loss)(q)

# Compatible with jit
jit_loss = jax.jit(loss)
l = jit_loss(q)

# Compatible with vmap
vmap_loss = jax.vmap(loss)
batch_q = jnp.tile(q, (10, 1))
batch_ls = vmap_loss(batch_q)

# Explicit warm starting
def step(state, q_k):
    sol = elastiqp.jax.solve(Q, q_k, G, h, penalty, A=A, b=b, warm_start=state)
    return sol, sol.x
init = elastiqp.jax.solve(Q, qs[0], G, h, penalty, A=A, b=b)
_, xs = jax.lax.scan(step, init, qs)

For runnable Python/JAX/PyTorch examples, see the examples folder

Benchmarks

Warning

Under development

See StanfordASL/elastiqp_benchmarks

Acknowledgments

ElastiQP builds on the following excellent projects:

• qpax
• DAQP
• ProxQP
• PIQP

ElastiQP is licensed under Apache 2.0; see LICENSE and NOTICE for third-party notices.

Citation

@article{morton2026elastiqp,
  author={Morton, Daniel and Arrizabalaga, Jon and Manchester, Zachary and Pavone, Marco},
  title={Elasti{QP}: An Always-Feasible QP Solver for Constrained Robot Control},
  journal={arXiv preprint arXiv:2609.19080},
  year={2026},
}