Research › Analysis

Agilicious: Open-Source and Open-Hardware Agile Quadrotor for Vision-Based Flight

Foehn, P. et al. · Science Robotics 7, eabl6259 (2022) · Paper

Q2 Computing analysis. All mathematical results and empirical findings are attributed to the original authors. We present our reading of the work as it relates to robotic automation.

What this paper solves

Agile quadrotor research requires hardware and software that is both open, reproducible, and capable enough to reach the physical limits of the platform. Proprietary or one-off platforms prevent the community from building cumulative results on a shared baseline. Agilicious provides a fully open-source, open-hardware quadrotor platform designed specifically for vision-based agile flight research, with a software stack capable of 5g maneuvers and a 1 kHz control loop.

Key mathematical framework

Quadrotor model and input parameterization

Agilicious uses a standard rigid-body quadrotor model. The state is $x = (p, v, q, ω) \in R^{3} \times R^{3} \times S^{3} \times R^{3}$ . The control input is collective thrust $c \in R$ and body angular acceleration $\overset{ω}{˙}^{B} \in R^{3}$ , mapped to individual motor thrusts ${f_{i}}_{i = 1}^{4}$ by the allocation matrix:

c τ_{x} τ_{y} τ_{z} = 1 - l l - κ 1 l - l - κ 1 l l κ 1 - l - l κ f_{1} f_{2} f_{3} f_{4}

where $l$ is the arm length and $κ$ is the drag-to-thrust coefficient. Individual motor thrusts are then mapped to motor speed commands.

Nonlinear MPC formulation

The primary controller is a nonlinear MPC that minimizes tracking error subject to dynamic feasibility. At each timestep, it solves:

u_{0 : N} min k = 0 \sum N ℓ (x_{k}, u_{k}, x_{k}^{ref}) s.t. x_{k + 1} = f (x_{k}, u_{k})

ℓ (x, u, x^{r}) = ∥ p - p^{r} ∥_{Q_{p}}^{2} + ∥ v - v^{r} ∥_{Q_{v}}^{2} + ∥ q ⊖ q^{r} ∥_{Q_{q}}^{2} + ∥ u ∥_{R}^{2}

The rotation error $q ⊖ q^{r}$ uses the quaternion log map to compute a rotation vector representing the angle between current and reference attitude. The solver runs at 50 Hz with a prediction horizon of 0.5 s, while an inner loop runs at 1 kHz for attitude stabilization.

Geometric control fallback

For compatibility with learned controllers that output different command formats, Agilicious also implements a geometric controller on $S O (3)$ . Given a desired thrust vector $f_{des}$ and heading, it constructs the desired rotation matrix and computes:

τ = - k_{R} e_{R} - k_{ω} e_{ω} + ω \times J ω

where $e_{R} = \frac{1}{2} (R_{des}^{⊤} R - R^{⊤} R_{des})^{\lor}$ is the rotation error on $S O (3)$ and $J$ is the inertia tensor.

Empirical results

Agility: Demonstrated 5g maneuvers including Power Loop and Barrel Roll on the physical platform
Control rate: 1 kHz attitude loop, 50 Hz MPC outer loop, with latency below 2 ms for the inner loop
Open hardware: Complete BOM, PCB files, and CAD released. Platform reproducible for under $1,000 in parts
Software compatibility: Supports ROS, Python, C++, and direct controller interfaces for learned policy deployment

What this means for robotic automation

Agilicious defines the open-hardware reference architecture for agile quadrotor research. Its motor allocation matrix, MPC formulation, and geometric controller represent the engineering choices that the field has converged on for platforms targeting 3-5g maneuvers. The 1 kHz control rate sets the latency ceiling that sensing and perception systems must achieve to feed the controller without becoming the bottleneck.

The open-source release is significant because it means the complete software stack used in AlphaPilot, Champion-Level Drone Racing, and Deep Drone Acrobatics can be reproduced, benchmarked, and extended on a known hardware baseline. For systems designed for edge compute, the Agilicious platform provides the physical deployment reference: any algorithm that runs on Agilicious's onboard compute (Jetson AGX Xavier) can be validated against a known performance envelope.