WarpMPC: Large-Batch MPC on GPU via ADMM with Unrolled LDL Factorization
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: We're looking at the paper: "WarpMPC: Large-Batch MPC on GPU via ADMM with Unrolled LDL Factorization".
Dev: Here's the gist of the paper. WarpMPC introduces numerical optimizations for maximizing throughput on GPU when solving large batches of sequential quadratic programming (SQP) iterations where all problems have the same structure.
Rosa: Problem Context and Motivation Model-predictive control (MPC) is a popular method for controlling nonlinear systems subject to constraints.
Dev: Recently, there has been significant interest in integrating deep learning and MPC for robotics, which requires solving large batches of MPC problems.
Rosa: Most existing MPC toolboxes are optimized for single solve speed on CPU or GPU or if batched on GPU can not handle general constraints.
Paper summary: Rosa: So we're looking at this paper, "WarpMPC: Large-Batch MPC on GPU via ADMM with Unrolled LDL Factorization," and the main idea is that they've developed some numerical optimizations for when you have a really big batch of sequential quadratic programming iterations running on a GPU.
Dev: The core thesis is that if all those problems in your batch share the same structure—meaning they have the same sparsity in time, cost, and constraints—you can speed things up by unrolling sparse linear factorizations and solves that are part of the alternating direction method of multipliers solver.
Taro: So what does this mean for us? It suggests that instead of solving each problem individually one after another, we can exploit that shared structure across the whole batch to avoid wasting computation and memory access bottlenecks.
Rosa: Exactly. They propose three specific optimizations: an optimized memory layout, a padding-reducing way to segment that unrolled factorization, and dependency level scheduled backsolves. They also accelerate sensitivity computation.
Dev: And the results they show are pretty impressive for throughput; they achieved speeds of eight thousand to two hundred fifty thousand sequential quadratic programming iterations per second on benchmarks like the nonlinear cartpole, quadrotor, and humanoid robot.
Taro: Outperforming existing methods by three times to twenty-five times is a significant jump if you're dealing with those kinds of large batches. It's not just theoretical; they even show practical relevance by distilling a neural network approximation from an MPC in under five minutes that stabilizes a nano quadrotor in hardware experiments.
Rosa: That stabilization part is really interesting for me, because it moves this from just math on the screen to something that actually works outside the lab environment. So, we've seen how they claim it can handle large batches of sequential quadratic programming iterations on GPUs via ADMM with unrolled LDL factorization >
Conclusion: Dev: To wrap up, we're talking about this paper "WarpMPC: Large-Batch MPC on GPU via ADMM with Unrolled LDL Factorization" by Hose, Jeon, Khazoom, Kim, and Trimpe. The authors are showing how to tackle the throughput challenge when you need to solve a large batch of sequential quadratic programming iterations on a GPU.
Rosa: It boils down to them finding that by unrolling sparse linear factorizations and solves related to the alternating direction method of multipliers solver, they can significantly boost performance by avoiding memory access issues and wasting computation.
Taro: From what I'm hearing, this work suggests that for problems where you have a lot of sequential MPC tasks with the same underlying structure, this approach using unrolled LDL factorization is much faster than current state-of-the-art GPU libraries.
Dev: The implication is that we can handle larger batches of constrained nonlinear model predictive control problems on GPUs with much higher speeds. They show consistent throughput improvements ranging from three times up to twenty-five times compared to other top GPU solvers for large batches >
Rosa: It’s about taking that idea and making it practical, proving it works on real robot benchmarks like the cartpole and quadrotor, and even showing how it can be used in a fast distillation process for an MPC policy >
Henrik Hose, Se Hwan Jeon, Charles Khazoom, Sangbae Kim, Sebastian Trimpe
Institute for Data Science in Mechanical Engineering, RWTH Aachen University, Germany · Massachusetts Institute of Technology, Cambridge MA, USA
cs.RO, cs.SY, eess.SY, math.OC
Submitted: 2026-07-13
Updated: 2026-11-16
Code: https://github.com/hshose/warpmpc
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 97/100
The gist: WarpMPC introduces numerical optimizations for maximizing throughput on GPU when solving large batches of sequential quadratic programming (SQP) iterations where all problems have the same structure.
Key concepts
- Sequential Quadratic Programming (SQP)
- SQP is a method used to find optimal solutions for complex constrained optimization problems, like those in MPC. It works by iteratively solving a sequence of simpler quadratic subproblems that approximate the original nonlinear problem. In this context, it's the core solver being accelerated.
- ADMM Solver
- Alternating Direction Method of Multipliers (ADMM) is an optimization algorithm used to solve large-scale problems by breaking them down into smaller, more manageable subproblems. WarpMPC uses ADMM as a framework to manage and efficiently solve the large batch of MPC problems in parallel.
- Unrolled Sparse LDL Factorization
- This technique involves pre-calculating sparse linear factorizations based on the shared structure (sparsity) found across many MPC problems. By 'unrolling' these factorizations, WarpMPC avoids redundant computations and memory access bottlenecks during the online solving phase, leading to massive speed improvements.
- Model Predictive Control (MPC)
- MPC is a control strategy for nonlinear systems that uses a model to predict future system behavior and then optimizes a sequence of control actions over a future time horizon. It is crucial for robotics, allowing systems like quadrotors or humanoid robots to navigate while respecting physical constraints.
Terminology
Summary
WarpMPC introduces numerical optimizations for maximizing throughput on GPU when solving large batches of sequential quadratic programming (SQP) iterations where all problems have the same structure. p. 0
Problem Context and Motivation
Model-predictive control (MPC) is a popular method for controlling nonlinear systems subject to constraints. Recently, there has been significant interest in integrating deep learning and MPC for robotics, which requires solving large batches of MPC problems. p. 0 Most existing MPC toolboxes are optimized for single solve speed on CPU or GPU or if batched on GPU can not handle general constraints. p. 0
Solver Architecture and Core Proposal
WarpMPC is a JAX toolbox with a Warp backend for solving large batches of MPC problems in parallel using SQP and a filter line search. The core idea is to exploit the shared sparsity in time, cost, and constraints across all problems in the batch by unrolling sparse linear factorizations and solves which dominate the alternating direction method of multipliers (ADMM) solver runtime. p. 0
Optimization Techniques for Throughput
The paper proposes several optimizations to avoid memory access bottlenecks and wasting computations. These optimizations include:
-
Optimized memory layout. p. 0
-
Padding-reducing segmentation of the unrolled factorization. p. 0
-
Dependency level scheduled backsolves. p. 0
Performance Benchmarks and Results
The optimizations achieve throughputs of 8 000 to 250 000 SQP iterations per second on nonlinear cartpole, quadrotor, and humanoid robot benchmarks. The practical usefulness is illustrated by synthesizing a dataset and training a neural network approximation of an MPC in under 4 minutes that stabilizes a nano quadrotor in hardware experiments. p. 0
Comparison with Existing Solvers
WarpMPC is compared against state-of-the-art libraries on GPU showing consistent throughput improvements of 3× to 25× for large batches. In nonlinear MPC examples, WarpMPC consistently improves throughput compared to sparse PCG (■). p. 6
Conclusion and Future Work
The paper concludes that WarpMPC significantly improves throughput by up to 24×. The toolbox is an open-source JAX toolbox for solving large batches of constrained nonlinear MPC problems on GPU via SQP and ADMM. p. 6 Future work may include Hessian regularization and GPU native interfaces for neural network models in the MPC constraints. p. 6
Index Terms
Optimization and Optimal Control; Software Tools for Robot Programming; Machine Learning for Robot Control.
Related Work Summary
Many CPU based MPC toolboxes exist, some support batched solves, but direct integration with GPU machine learning frameworks is limited.
Table Summary
Table I summarizes various GPUs solvers for trajectory optimization and MPC, showing WarpMPC's SQP with ADMM Unrolled sparse LDL⊤ implementation is superior in several metrics.
Algorithm Summary
Algorithm 1 outlines the offline phase (CPU/JIT setup) and the online phase (GPU batched execution) for WarpMPC. The process involves symbolic factorization for fixed-pattern K bars and then an online phase where each SQP iteration is solved using the ADMM update scheme. p. 3
Illustrative Example Summary
The paper illustrates practical relevance by distilling a neural network approximation of an MPC in under 5 min that stabilizes a nano quadrotor in hardware experiments. This is achieved by synthesizing large datasets and training policies on a single GPU within less than 4 min. p. 6
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
-
Large-batch MPC on GPU via ADMM with Unrolled LDLtop Factorization allows for solving batches of 10,000 to over 100,000 sequential quadratic programming iterations per second on nonlinear cartpole, quadrotor, and humanoid robot benchmarks.
-
The system can achieve throughputs of 8,535 SQP iterations per second on the humanoid benchmark using optimal column segmentation and level-scheduled triangular solves.
-
The AI system can synthesize a neural network approximation of an MPC for embedded hardware deployment onboard a nano quadrotor in under 4 minutes from a large dataset of 20 million state-action pairs.
-
The system can perform forward solve and backward sensitivities for QP problems at throughputs up to 279,000 solves per second without sensitivities or 9,570 solves per second with sensitivities on linear MPC benchmarks.
-
The AI system can stabilize a nano quadrotor in hardware experiments by deploying a neural network approximation of an MPC trained within less than 4 minutes.
Sources
- Residual MPC: Blending Reinforcement Learning with GPU-Parallelized Model Predictive Control
- Exploiting GPU/SIMD Architectures for Solving Linear-Quadratic MPC Problems
- Fast and Certifiable Trajectory Optimization
- GATO: GPU-Accelerated and Batched Trajectory Optimization for Scalable Edge Model Predictive Control
- TurboMPC: Fast, Scalable, and Differentiable Model Predictive Control on the GPU
- On the Differentiability of the Primal-Dual Interior-Point Method
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