Tiny Learning-Based MPC for Multirotors: Solver-Aware Learning for Efficient Embedded Predictive Control

TL;DR

This work tackles high-rate, constraint-aware control for tiny multirotors under nonlinear dynamics and model mismatch by marrying differential flatness with Gaussian Process-based disturbance learning in a co-designed MPC framework. The key innovation, LinGP, yields an affine mean and diagonal quadratic covariance, enabling probabilistic tightening to SOC constraints; a purpose-built ADMM solver runs on an embedded microcontroller to achieve $100$ Hz control. The authors demonstrate onboard deployment on a $53$ g Crazyflie 2.1 platform, achieving up to a $43\%$ gain in tracking accuracy over existing embedded MPC methods under disturbance, and achieving the first LB MPC onboard on such a small aerial vehicle. The approach offers practical impact for robust, high-rate predictive control in resource-constrained drones, with applications in environmental monitoring, search-and-rescue, and other field tasks.

Abstract

Tiny aerial robots hold great promise for applications such as environmental monitoring and search-and-rescue, yet face significant control challenges due to limited onboard computing power and nonlinear dynamics. Model Predictive Control (MPC) enables agile trajectory tracking and constraint handling but depends on an accurate dynamics model. While existing Learning-Based (LB) MPC methods, such as Gaussian Process (GP) MPC, enhance performance by learning residual dynamics, their high computational cost restricts onboard deployment on tiny robots. This paper introduces Tiny LB MPC, a co-designed MPC framework and optimization solver for resource-constrained micro multirotor platforms. The proposed approach achieves 100 Hz control on a Crazyflie 2.1 equipped with a Teensy 4.0 microcontroller, demonstrating a 43% average improvement in tracking performance over existing embedded MPC methods under model uncertainty, and achieving the first onboard implementation of LB MPC on a 53 g multirotor.

Figures (8)

• Figure 1: Hardware setup for the proposed Tiny Learning-Based MPC framework. The controller runs entirely onboard a (1) Teensy 4.0 microcontroller (600 MHz, 1 MB RAM) mounted on a (2) custom expansion board and interfaced with a (3) Crazyflie 2.1 platform. The complete system, including (4) upgraded motors and propellers (maximum collective thrust 0.764 N), weighs 53 g in total. The setup demonstrates full onboard learning-based predictive control on a resource-constrained multirotor.
• Figure 2: Block diagram of the proposed Tiny LB MPC framework. The controller integrates the multirotor’s differential flatness, a Linearized Gaussian Process (LinGP) model of residual dynamics, and a linear MPC with second-order cone (SOC) constraints to compute the control input $\mathbf{u}$ at 100 Hz on the Teensy 4.0 microcontroller.
• Figure 3: Simulation 1 — Effect of GP training data size on RMSE for a 3 rad/s circular trajectory over 20 trials. The solver runs 5 ADMM iterations; 20 data points achieve low RMSE with minimal variance.
• Figure 4: Simulation 2 — Figure-8 trajectories tracked by Tiny FB MPC (yellow) and Tiny LB MPC (blue) at increasing angular speeds. Tiny LB MPC achieves lower RMSEs: 0.032 vs. 0.10 m (0.5 rad/s), 0.086 vs. 0.152 m (1 rad/s), and 0.140 vs. 0.193 m (right).
• Figure 5: Simulation 3 — Thrust-angle violations ($\theta_{\max}=25^{\circ}$) across trajectory frequencies and ADMM iterations. Tiny LB MPC remains feasible with under five iterations, while higher frequencies reveal a trade-off between solver speed and constraint satisfaction.