2026-08-13 –, Room 6
We present recent developments in the control-toolbox ecosystem, centered around OptimalControl.jl v2.0.0. This Julia-based environment enables modeling and solving optimal control problems through a mathematical domain-specific language, supporting direct methods on CPU and GPU. We introduce OptimalControlProblems.jl, a benchmark collection of optimal control problems, and CTBenchmarks.jl, a systematic benchmarking framework comparing various modelers (JuMP, ADNLPModels, ExaModels) and solvers (Ipopt, MadNLP, MadNCL) on CPU and GPU.
OptimalControl.jl provides a high-level, mathematical DSL for modeling optimal control problems that combines ease of use with computational performance. The package supports direct transcription methods (discretization with Runge-Kutta schemes), running efficiently on both CPU and GPU architectures. This makes it a dedicated environment for optimal control that bridges the gap between mathematical formulation and efficient numerical computation.
The upcoming v2.0.0 release brings significant enhancements to the ecosystem. To systematically evaluate performance across different computational approaches, we have developed two new packages. OptimalControlProblems.jl [1] provides a curated collection of optimal control problems modeled with both JuMP and OptimalControl, designed specifically for benchmarking purposes. CTBenchmarks.jl [2] implements a flexible framework for comparing various combinations of modelers (JuMP [3], ADNLPModels [4], ExaModels [5]) and solvers (Ipopt [6], MadNLP [7], MadNCL [8]), with an extensible architecture that facilitates adding new solvers such as Uno [9]. Our initial benchmarks reveal interesting performance characteristics on both CPU and GPU, demonstrating that specialized optimal control formulations can outperform general-purpose modeling frameworks in certain scenarios.
The talk will present the new features of OptimalControl.jl v2.0.0, the benchmarking framework, and CPU/GPU performance results.
References:
[1] OptimalControlProblems.jl
[2] CTBenchmarks.jl
[3] JuMP.jl
[4] ADNLPModels.jl
[5] ExaModels.jl
[6] NLPModelsIpopt.jl
[7] MadNLP.jl
[8] MadNCL.jl
[9] Uno
[10] control-toolbox.org
Associate professor, applied mathematics
Université de Toulouse, INP-ENSEEIHT & IRIT, CNRS, France