WAND: Learning Robust Navigation under Complex Wind Disturbances and Dense Obstacles for Quadrotors

arXiv:2610.11809 · cs.RO · Submitted 2026-10-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "WAND: Learning Robust Navigation under Complex Wind Disturbances and Dense Obstacles for Quadrotors".

Dev: The gist Robust navigation in cluttered environments remains a fundamental challenge for quadrotors, particularly when strong wind disturbances arise, which perturb vehicle dynamics, limit control authority,

Rosa: First, who's behind it and why it matters.

Paper summary: Rosa: So, we’re looking at this paper called "WAND: Learning Robust Navigation under Complex Wind Disturbances and Dense Obstacles for Quadrotors." The core idea here is tackling that fundamental problem with quadrotors when strong winds hit them in crowded spaces.

Dev: Right. Basically, robust navigation in cluttered environments is tough for these machines, especially when a lot of wind disturbs the dynamics and makes controlling it harder while increasing the risk of crashes.

Rosa: This WAND framework proposes using reinforcement learning to handle those time-varying wind disturbances in dense obstacle fields. The main claim is that it estimates this disturbance acceleration directly from past proprioceptive states using a Temporal Convolutional Network, or TCN.

Dev: And they integrate that estimate into the policy through something called a zero-initialized residual module called the WindAdapter, which also gives you feedforward compensation for the low-level control commands. So it’s doing two things at once: condition-aware navigation and feedforward rejection of those disturbances.

Taro: What I find interesting is how they are coupling that disturbance estimation with the actual policy, which means the system uses this estimate both to guide its navigation decisions and to pre-compensate for what the wind is doing right now. It’s a dual use of the estimate, which makes it quite clever.

Rosa: Exactly. They define their disturbance as the difference between what you see from your rotors under relative airflow and the standard collective thrust command, which feeds into an input for that TCN estimator. That input window is defined by states like normalized collective thrust commands and hover-thrust estimates.

Dev: And the TCN takes that historical window of inputs, which they call Xt =

xt−L+one <ref:2610.11809#pg1>..., xt: , and it maps that history to an acceleration estimate, denoted as aˆd,W t. This output is in three dimensions.

Taro: So they are essentially using the recent history of what the drone has been doing—its thrust commands and hover estimates—to predict how fast the wind is pushing it right now, which is a pretty direct way to infer effects that are otherwise implicit.

Rosa: That's the gist of their disturbance estimation. Then, this estimate goes into an augmented observation ot for the policy network, where it’s summarized by that TCN output and fused with obstacle features.

Dev: The policy then takes this augmented observation and selects an action ut based on a network π. They also transform that world-frame disturbance estimate into a frame h before encoding it into a wind-conditioned feature zwind via the WindAdapter.

Taro: And that WindAdapter module is where the magic happens, taking the disturbance estimate and turning it into a feature that conditions the policy input, which is then added elementwise to the proprioceptive state features to make z.

Rosa: So they’re combining obstacle perception and current states with this wind-conditioned feature for navigation decisions. The actor outputs a desired acceleration command ut in the heading frame, and that gets combined with the feedforward compensation term to create the final world-frame setpoint a cmd,W t.

Dev: And they define their reward structure around five terms: velocity, obstacle proximity, height error, wind effect itself for some feedback, and a term rterm. One specific part of that is rwind, which measures directional clearance based on the disturbance estimate and the desired acceleration.

Taro: So it’s not just about reaching a point; they are explicitly rewarding how well you navigate *given* the wind condition, even by penalizing you for being in a spot where the disturbance is pushing you in a bad direction.

Rosa: Overall, this paper shows that combining an explicit estimator with the policy network and that residual module makes WAND quite robust. We’ll talk about what this means for real-world flight next.

Conclusion: Rosa: So, to wrap up, we’re looking at WAND by Tang, Li, Liang, You, Li, Zhao, Wang and Li. The title itself points to what they did: learning robust navigation in cluttered environments under complex wind disturbances.

Dev: It means the system can handle those strong wind disturbances that mess with vehicle dynamics and limit control authority without crashing easily when there are lots of obstacles around. It's about inferring the disturbance effects implicitly, which is where previous learning policies often struggled with partial observability.

Rosa: The key contribution here is proposing WAND, which explicitly estimates the wind-induced disturbance acceleration from historical proprioceptive states using a TCN and integrating it via that WindAdapter residual module. They also reuse that estimate for feedforward compensation during execution, which is a smart way to handle the uncertainty.

Taro: What this really means for autonomous systems is moving beyond just relying on perception and proprioception to infer things implicitly when the environment changes rapidly, like a sudden gust of wind. It gives you an explicit model of the disturbance that you can use to actively counteract it in real time.

Dev: The experimental validation shows that WAND improved success rates by eight point three percentage points on average compared to just using feedforward compensation alone across twelve different simulation settings <ref:2610.11809#pg3>. Plus, they showed wind-direction-dependent trajectory adaptation in controlled opposite crosswind tests and succeeded in eighteen out of twenty indoor fan-induced flight tests.

Rosa: So, for someone who only listens to the show, WAND means that a quadrotor can maintain much better navigation accuracy when it’s flying indoors or outdoors where the wind isn't constant or predictable. It makes the system more reliable in messy situations.

Taro: The limitation they point out is that they are focusing on time-varying disturbances inferred from proprioceptive histories, and they mention that systematic evaluation under spatially varying wind fields is still future work, meaning we don’t have a full picture of how it handles every kind of complex spatial variation yet.

Dev: So while it works really well in controlled settings, the next step is testing it when the wind changes in complicated ways across different areas, not just during one specific test run. That’s where the real scaling challenge lies for control engineers.

Rosa: Exactly. So WAND gives us a much more robust navigation tool for these drones, moving them from lab settings to genuinely cluttered and windy real-world applications, provided we can handle those spatial variations in the future testing.

University of Science and Technology of China

cs.RO

Submitted: 2026-10-08

Updated: 2026-10-08

Journal ref: IEEE Robotics and Automation Letters, vol. 11, no. 11, pp. 12392-12399, Nov. 2026

DOI: 10.1109/LRA.2026.3730214

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: The gist Robust navigation in cluttered environments remains a fundamental challenge for quadrotors, particularly when strong wind disturbances arise, which perturb vehicle dynamics, limit control

Key concepts

Wind Disturbance Estimation
This process uses a Temporal Convolutional Network (TCN) to predict the acceleration caused by wind. It looks at historical flight data—like thrust commands and hover thrust estimates—to calculate the current aerodynamic disturbance acceleration, allowing the system to anticipate wind effects before they significantly impact flight.
WindAdapter
This is a zero-initialized residual module integrated into the policy network. It takes the estimated wind disturbance information and transforms it into a feature that conditions the navigation decision. This allows the learned policy to adapt its movement strategy based on real-time wind conditions.
Feedforward Compensation
This technique uses the estimated disturbance acceleration ($\hat{a}_{d,h}$) to calculate a corrective term that is added directly to the desired control command. This provides immediate, low-level rejection of the estimated wind forces during execution, ensuring better tracking and stability.
Wind-Aware Policy Network
The policy network takes combined observations—including obstacle features and the TCN's disturbance estimate—to decide on an action. It fuses proprioceptive data with the wind-conditioned feature to select a movement command that is specifically tailored to compensate for the current wind environment.

Terminology

Summary

The gist Robust navigation in cluttered environments remains a fundamental challenge for quadrotors, particularly when strong wind disturbances arise, which perturb vehicle dynamics, limit control authority, and substantially increase collision risk.

WAND Framework Overview

The paper proposes WAND (Wind-Aware Navigation with Disturbance Estimation), a reinforcement learning framework for navigation under time-varying wind disturbances in dense obstacle fields. Specifically, WAND estimates wind-induced disturbance acceleration from historical proprioceptive states using a Temporal Convolutional Network (TCN). This estimation is integrated into the policy via a zero-initialized residual module, WindAdapter, while simultaneously providing feedforward compensation for lowlevel control. The dual use of the estimate couples disturbance-conditioned navigation with feedforward disturbance rejection.

Wind Disturbance Estimation

The estimator targets translational disturbances by defining the aerodynamic disturbance as the difference between the total rotor force under relative airflow and the nominal collective thrust. The estimator input is defined as xt = Tt,(qˆt)T,(aˆB t)T,(vˆB t)T, where Tcmd t is the normalized collective-thrust command generated by PX4, and Thover is the hover-thrust value estimated online by the PX4 Hover Thrust Estimator on the same command scale. The TCN maps this causal window Xt = [xt−L+1,..., xt] ∈ R L×11 to the current disturbance acceleration, denoted as aˆd,W t = fTCN(Xt) ∈ R 3.

Wind-Aware Policy Network

The policy receives an augmented observation ot in which the TCN estimate summarizes the recent proprioceptive history, and selects an action ut ∼ π(· ot). The observation ot comprises proprioceptive, exteroceptive, and estimated-disturbance information. The state fusion encoder processes the obstacle feature, which is concatenated with Sprop,t to produce znav. Before being provided to the policy, the world-frame disturbance estimate is transformed into frame h as aˆd,h t = RhWaˆd,W t. The WindAdapter encodes aˆd,h t into a wind-conditioned feature zwind with the same dimensionality as znav. The two features are fused by elementwise addition to form the input to the policy network, z = znav + zwind.

Action and Reward Structure

The actor outputs the desired acceleration command ut = a des,h t ∈ [−3, 3]3 m/s 2 in the heading frame. This output is then combined with the feedforward disturbance compensation term for execution to yield the worldframe acceleration setpoint a cmd,W t = RWh a des,h t − aˆd,h t. The overall reward r is defined as a weighted sum of five terms: r = λ1rvel + λ2robs + λ3rheight + λ4rwind + rterm. The directional clearance used by rwind is defined as (i∗, j∗) = arg max (i,j) (eˆh i,j) T aˆd,h t∥aˆd,h t∥ > 0.1 m/s 2.

Experimental Validation

Across 12 wind-disturbed simulation settings, WAND improved the observed success rate by 8.3 percentage points on average relative to feedforward compensation alone. Controlled oppositecrosswind experiments further showed wind-direction-dependent trajectory adaptation. In indoor fan-induced flight tests, WAND succeeded in 18 of 20 trials, demonstrating the feasibility of realtime onboard navigation. The TCN achieves the lowest RMSE under all four conditions, including the held-out mixedwind condition. WAND obtains a higher observed success rate in all 12 wind-disturbed settings, with an average improvement of 8.3 percentage points. WAND succeeds in all trials under both directions when compared to Base+Comp.

Conclusion

By combining a navigation policy with a wind disturbance estimator and the proposed WindAdapter module, WAND enables the quadrotor to adapt its motion strategy according to the wind condition. Experiments in simulation and controlled indoor fan tests demonstrated improved navigation robustness. Systematic evaluation under spatially varying wind fields remains future work.

The gist Robust navigation in cluttered environments remains a fundamental challenge for quadrotors, particularly when strong wind disturbances arise, which perturb vehicle dynamics, limit control authority, and substantially increase collision risk.

How it works

WAND estimates wind-induced disturbance acceleration from historical proprioceptive states using a Temporal Convolutional Network (TCN). This estimation is integrated into the policy via a zero-initialized residual module, WindAdapter, while simultaneously providing feedforward compensation for lowlevel control. The dual use of the estimate couples disturbance-conditioned navigation with feedforward disturbance rejection.

Action and Reward Structure

The actor outputs the desired acceleration command ut = a des,h t ∈ [−3, 3]3 m/s 2 in the heading frame. This output is then combined with the feedforward disturbance compensation term for execution to yield the worldframe acceleration setpoint a cmd,W t = RWh a des,h t − aˆd,h t.

Improvements for AI systems

  1. textbfIncrease Robustness via Explicit Disturbance Estimation and Compensation: The WAND framework estimates wind-induced disturbance acceleration from historical proprioceptive states using a Temporal Convolutional Network (TCN) and integrates this into the policy via a zero-initialized residual module, WindAdapter. This dual use of the estimate couples disturbance conditioned navigation with feedforward disturbance rejection, leading to an average improvement of 8.3 percentage points in success rate over baseline methods.

  2. textbf Achieve Wind-Direction-Dependent Trajectory Adaptation: Controlled oppositecrosswind experiments showed wind-direction-dependent trajectory adaptation. WAND's policy combines these features with obstacle perception and the current proprioceptive state, allowing it to adapt its safety margin based on wind magnitude and direction, as evidenced by WAND succeeding in all trials under both crosswind directions compared to Base+Comp.

  3. textbf Enable Real-Time Onboard Navigation Under Cluttered Conditions: The system can perform realtime navigation in cluttered indoor environments under fan-induced airflow, successfully completing 18 out of 20 trials in indoor fan tests by using the WindAdapter to encode disturbance estimates into wind-conditioned features.

  4. textbf Enhance Disturbance Estimation Accuracy Through Causal History: The TCN estimator inputs a causal history of measurements, defined as Xt = [xt−L+1,..., xt] ∈ R L×11, where it maps this history to the current disturbance acceleration to capture the disturbance acting at the vehicle’s location from recent proprioceptive measurements.

  5. textbf Provide Dynamic Safety Margins Based on Disturbance: The reward function explicitly couples wind estimation with obstacle clearance via the term rwind = −h∥aˆd,h,t∥ · ReLU0.3∥aˆd,h,t∥ + 0.5 − dwind,t i2, which penalizes insufficient directional clearance and encourages a larger local safety margin as the estimated disturbance increases.

Sources

Related papers