HDVIO2.0: Wind and Disturbance Estimation with Hybrid Dynamics VIO (T-RO 2025)

Autonomous quadrotors rely on accurate

state estimation to navigate safely.

Visual inertial model odometry combines

camera and IMU data with the quadrotor's

model to estimate the full vehicle

state. The accuracy of current methods

degrades in the presence of strong

disturbances or when the dynamics are

not well modeled.

This factor graph shows how a standard

tightly coupled VIO system operates. It

fuses vision factors obtained through

feature tracking and inertial factors

obtained from the IMU to estimate the

quadrotor state along with the IMU

biases.

Visual inertial models extend the

traditional VIO framework by

incorporating a motion prior based on

quadrotor dynamics.

While these state-of-the-art systems

perform well in many scenarios, their

accuracy degrades in the presence of

inaccurate vehicle models or persistent

external disturbances such as wind due

to the simplified assumptions in the

dynamics model. In our method, we

address these limitations by proposing a

hybrid dynamics model that combines a

first principles quadrotor model with a

learningbased component that captures

residual effects such as aerodynamic

drag. It models the translational and

rotational vehicle dynamics and tightly

integrates them into the VIO system with

minimal runtime overhead. In contrast to

prior work, our learn dynamics model

does not require access to the full

drone state. Only the commanded thrust

and torqus as well as the gyroscope

measurements are needed. These inputs

are processed by two temporal

convolutional networks. One predicts a

residual thrust and the other predicts a

residual torque. The predicted residuals

are then added to the measured thrust

and torque and the resulting signals are

integrated to estimate updates in

velocity, position, and relative

orientation. We validate our learned

dynamics model with a neurobam data set

which contains very fast and agile

trajectories. Our method performs

remarkably well compared to modelbased,

learning based and hybrid baselines

despite not having access to the full

quadrotor state.

In terms of trajectory estimation, our

method outperforms the visual inertial

and visual inertial model baselines on

the Blackbird data set. The Blackbird

data set contains diverse trajectories

at medium speeds. Compared to the

baselines, the largest improvements are

in the faster trajectories where the

camera motion and rapid yaw changes make

the tracking of visual features

challenging.

Next, we move to real world experiments

to demonstrate the performance of our

hybrid dynamics VIO pipeline in the

presence of disturbances. We equip a

quadrotor with a dragboard and fly it in

strong winds up to 25 km/h.

On the bottom right, the force estimate

and on the left the position estimate

from the pipeline are shown.

Our method is plotted in red and the

ground truth positions are shown in

blue.

Despite the challenging flight

conditions, our method is very accurate.

Finally, we want to demonstrate that

HDVIO2 runs efficiently on board the

quadrotor and provides real-time state

estimates for closed loop control.

The goal of this experiment is to track

a random trajectory and only use our

HDVIO2 state estimate for control. The

plot at the bottom shows the ground

truth position in blue, the Intel Real

Sense estimate in yellow, and ours in

red.

Notably, HDVIO2 outperforms the

commercial stereo-based visual inertial

slam system of the Intel Real Sense T265

as shown in the plot. Towards the end of

the 120 seconds flight, the Real Sense

has drifted 50 cm, whereas our approach

only drifts about 20 cm.