Code Generation and Conic Constraints for Model-Predictive Control on Microcontrollers with Conic-TinyMPC

TL;DR

This work targets real-time model-predictive control on microcontrollers with conic constraints, a domain where conventional solvers struggle due to memory and compute limits. It extends TinyMPC to Conic-TinyMPC by adding SOCP support and high-level code-generation interfaces (Python, MATLAB, Julia), leveraging ADMM with cached precomputations and an infinite-horizon LQR approximation to drastically reduce online complexity and memory footprint. The approach achieves up to $142.7\times$ speedups and enables larger problems to fit on tiny MCUs, with extensive microcontroller benchmarks and hardware experiments on a Crazyflie quadrotor demonstrating predictive safety filtering, attitude-constrained regulation, and conic glide-slope tracking. The work provides open-source tooling and practical demonstrations, enabling safe, real-time embedded conic MPC for robotics and related constrained control tasks, with future directions toward nonlinear models via reachability-based disturbances.

Abstract

Model-predictive control (MPC) is a powerful framework for controlling dynamic systems under constraints, but it remains challenging to deploy on resource-constrained platforms, especially for problems involving conic constraints. To address this, we extend recent work developing fast, structure-exploiting, cached ADMM solvers for embedded applications, to provide support for second-order cones, as well as C++ code generation from Python, MATLAB, and Julia for easy deployment. Microcontroller benchmarks show that our solver provides up to a two-order-of-magnitude speedup, ranging from 10.6x to 142.7x, over state-of-the-art embedded solvers on QP and SOCP problems, and enables us to fit order-of-magnitude larger problems in memory. We validate our solver's deployed performance through simulation and hardware experiments, including conically-constrained trajectory tracking on a 27g Crazyflie quadrotor. To get started with Conic-TinyMPC, visit our documentation, examples, and the open-source codebase at https://tinympc.org.

Figures (4)

• Figure 1: We demonstrate our solver using a 27 gram nano quadrotor, the Crazyflie. Top: we track a descending helical reference (red) with its position subject to a 45$^{\circ}$ second-order cone glideslope. This requires the aircraft to perform a spiral landing maneuver (blue). Bottom: we design a predictive safety filter to guarantee safe maneuvers within a box-shaped space (blue) regardless of the nominal controller behavior (red).
• Figure 2: The tree structure of the generated code. The main program is stored in tinymain.cpp.
• Figure 3: (a) Predictive safety filtering performance comparison between Conic-TinyMPC and OSQP on an STM32F405 Feather board. Top row shows average iteration times, bottom row shows memory usage. Left column: time horizon kept constant at $N=10$ while state dimension $n$ ranged from 2 to 32 and input dimension was set to half of the state dimension. Right column: state and control input held constant at $n=10$ and $m=5$ while $N$ ranged from 4 to 100. Error bars represent maximum and minimum time taken per iteration for all MPC steps. Black dotted lines denote memory thresholds. (b) Rocket soft-landing performance comparison between Conic-TinyMPC, ECOS, and SCS using a Teensy 4.1 development board. Top plot shows memory usage, bottom plot shows average iteration times. In this SOCP-based experiment $n=6$ and $m=3$ while $N$ varied from 2 to 256. Error bars represent maximum and minimum time taken per iteration for all MPC steps performed. Black dotted lines denote memory thresholds.
• Figure 4: Attitude/thrust vector regulating performance of different controllers on the Crazyflie. While Conic-TinyMPC was able to constrain the aircraft attitude within the bounds (dashed lines for 0.25 and 0.2 radians), Brescianini and Mellinger exhibited large attitude deviations, causing failures.