Tiny Learning-Based MPC for Multirotors
Welcome to the Tiny Learning-Based MPC project! This page covers the progress made on this project to date by members of Robora Lab.
Background
Micro aerial vehicles (MAVs), such as quadrotors, are increasingly critical in applications requiring agile, autonomous navigation in constrained and dynamic environments—ranging from search-and-rescue to environmental monitoring. However, achieving robust and high-performance control on these platforms is challenging due to two fundamental limitations: the effects of unmodeled aerodynamic disturbances at high speeds, and the severe computational constraints of onboard processors, especially in sub-100 g multirotors like the Crazyflie 2.1.
The research in this project addresses both challenges by developing learning-based Model Predictive Control (MPC) algorithms that are not only aware of the complex, nonlinear aerodynamics affecting multirotor flight, but also designed to run in real time on embedded microcontrollers. Traditional flatness-based planning and control methods for multirotors often assume simplified dynamics, neglecting the nonlinear and input-dependent effects of aerodynamic forces such as drag. These simplifications result in poor tracking performance and feasibility issues, particularly during aggressive maneuvers or in windy environments.
To overcome this, we learn the mapping from flat outputs to multirotor thrust using Gaussian Processes (GPs), enabling the controller to anticipate and adapt to aerodynamic disturbances. This GP-enhanced representation allows us to formulate a convex MPC problem, efficiently solved as a second-order cone program (SOCP), while retaining guarantees on physical feasibility. Moreover, to meet the tight resource constraints of small platforms, we design solver-aware implementations of learning-based MPC that achieve control rates of up to 100 Hz on a 53 g Crazyflie with a Teensy 4.0 processor.
Contributions of the project are listed below from newest to oldest.
Tiny Learning-Based MPC for Multirotors: Solver-Aware Learning for Efficient Embedded Predictive Control [Summer 2024]
Submmited to RA-L. Check out the paper here!
Tiny aerial robots show promise for applications like environmental monitoring and search-and-rescue but face challenges in control due to their limited computing power and complex dynamics. Model Predictive Control (MPC) can achieve agile trajectory tracking and handle constraints. Al- though current learning-based MPC methods, such as Gaussian Process (GP) MPC, improve control performance by learn- ing residual dynamics, they are computationally demanding, limiting their onboard application on tiny robots. This paper introduces Tiny Learning-Based Model Predictive Control (LB MPC), a novel framework for resource-constrained micro mul- tirotor platforms. By exploiting multirotor dynamics’ structure and developing an efficient solver, our approach enables high- rate control at 100 Hz on a Crazyflie 2.1 with a Teensy 4.0 microcontroller. We demonstrate a 23% average improvement in tracking performance over existing embedded MPC methods, achieving the first onboard implementation of learning-based MPC on a tiny multirotor (53 g).
Tiy L MPC
Tiny FB MPC.mp4
Tiny LB MPC.mp4
A Computationally Efficient Learning-Based Model Predictive Control for Multirotors under Aerodynamic Disturbances [Winter 2024]
Accepted as a contributing paper at the 2024 International Conference on Unmanned Aircraft Systems (ICUAS). Check out the paper here!
Neglecting complex aerodynamic effects hinders high-speed yet high-precision multirotor autonomy. In this paper, we present a computationally efficient learning-based model predictive controller that simultaneously optimizes a trajectory that can be tracked within the physical limits (on thrust and orientation) of the multirotor system despite unknown aerodynamic forces and adapts the control input. To do this, we leverage the well-known differential flatness property of multirotors, which allows us to transform their nonlinear dynamics into a linear model. The main limitation of current flatness-based planning and control approaches is that they often neglect dynamic feasibility. This is because these constraints are nonlinear as a result of the mapping between the input, i.e., multirotor thrust, and the flat state. In our approach, we learn a novel representation of the drag forces by learning the mapping from the flat state to the multirotor thrust vector (in a world frame) as a Gaussian Process (GP). Our proposed approach leverages the properties of GPs to develop a convex optimal controller that can be iteratively solved as a second-order cone program (SOCP). In simulation experiments, our proposed approach outperforms related model predictive controllers that do not account for aerodynamic effects on trajectory feasibility, leading to a reduction of up to 55% in absolute tracking error.
SOCP No Learning
SOCP Learning