PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Zhonghan Tang, Chenhui Li, Shuai Liang, Zhongrui You, Jianan Li, Bin Zhao, Zhigang Wang, Xuelong Li

arXiv:2610.11809v1cs.RO

TL;DR

Wind-aware navigation is challenging because policies relying on obstacle perception and proprioception must infer time-varying wind effects implicitly. WAND estimates disturbance acceleration from proprioceptive history and uses the estimate for policy conditioning and feedforward compensation. Across 12 wind-disturbed simulation settings, it improved observed success rate by 8.3 percentage points on average relative to Base+Comp.

  • Problem

    Learning-based navigation policies typically condition on obstacle perception and proprioception rather than explicit wind estimates, leaving time-varying aerodynamic effects implicit under partial observability.

    Source: §I · p. 2

  • Method

    WAND estimates wind-induced disturbance acceleration from proprioceptive history and uses it to condition the navigation policy and provide feedforward compensation.

    Source: §I · p. 2

  • Results

    8.3 percentage points average observed success-rate improvement over Base+Comp across 12 wind-disturbed simulation settings.

    Source: B. Simulation Results · p. 6

  • Takeaways & Limitations

    WAND adapts its motion strategy according to wind conditions, with simulation and controlled indoor fan tests demonstrating improved navigation robustness.

    Source: §V · p. 7

    Systematic evaluation under spatially varying wind fields remains future work.

    Source: §V · p. 7

Abstract

from arXiv · show

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. Existing learning-based navigation policies typically rely on obstacle perception and proprioceptive observations, requiring the policy to infer time-varying disturbance effects implicitly and thereby limiting robustness under partial observability. This 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, \emph{WindAdapter}, while simultaneously providing feedforward compensation for low-level control. The dual use of the estimate couples disturbance-conditioned navigation with feedforward disturbance rejection. 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 opposite-crosswind 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 real-time onboard navigation.

I. INTRODUCTION

Quadrotors navigating clutter face time-varying wind that can increase collision risk, while learning policies typically infer wind effects implicitly from obstacle and proprioceptive observations. WAND estimates disturbance acceleration from proprioceptive history and uses it for both policy conditioning and feedforward compensation.

I. INTRODUCTION

  • Time-varying airflow in confined environments and limited actuation authority make micro quadrotors susceptible to wind, increasing collision risk in clutter.Traditional approaches may leave residual tracking errors under strong, rapidly varying wind, eroding trajectory clearance.Source: §I · p. 1
  • Learning-based navigation policies primarily use obstacle perception and proprioception, so they infer changing aerodynamic effects implicitly despite identical observations sometimes requiring different actions under different winds.This partial observability complicates robust wind-aware navigation.Source: §I · pp. 1–2
  • WAND estimates wind-induced disturbance acceleration from proprioceptive histories, encodes it through a zero-initialized residual WindAdapter for policy conditioning, and reuses the estimate for feedforward compensation.The policy also combines the wind-conditioned features with obstacle perception and the current proprioceptive state.Source: §I · p. 2
  • WAND is validated through simulation and real-world flight experiments under diverse wind conditions, with improved robustness and disturbance-adaptive behavior compared with baseline methods.The introduction illustrates onboard deployment in a cluttered indoor scene under fan-induced airflow.Source: Figure 1 · p. 1, §I · p. 2

A. Hierarchical Navigation in Cluttered Environments

Hierarchical planners emphasize geometric feasibility and rely on controllers to track their trajectories, while learning-based navigation integrates perception and control. WAND combines explicit wind-disturbance estimation with obstacle-aware policy adaptation and low-level compensation.

A. Hierarchical Navigation in Cluttered Environments

  • Hierarchical planners such as Fast-Planner and EGO-Planner generate collision-free trajectories from occupancy-map geometry, but typically rely on low-level controllers for accurate tracking.Disturbance rejection is often left to tracking control, so severe gusts can drive a vehicle beyond intended safety margins.Source: A. Hierarchical Navigation in… · p. 2

B. Learning-Based Navigation

  • Learning-based methods learn obstacle avoidance and goal reaching, with examples using sim-to-real transfer, LiDAR representations, dynamic-obstacle policies, and memory-augmented latent representations.The cited examples include CAD2RL, NavRL, and Yu et al..Source: B. Learning-Based Navigation · p. 2
  • Unlike history-conditioned latent adaptation, WAND conditions navigation on a supervised, physically interpretable estimate of wind-induced disturbance acceleration and reuses it for feedforward compensation.This connects explicit onboard disturbance estimation – with adaptive learning and control under wind,.Source: B. Learning-Based Navigation · p. 2

III. METHODOLOGY

WAND’s framework combines obstacle perception, temporal wind-disturbance estimation, policy conditioning, and feedforward compensation.

III. METHODOLOGY

  • The framework uses ray-cast obstacle features and a TCN estimate from proprioceptive history to condition the policy, then combines its desired acceleration command with feedforward disturbance compensation.The policy receives the disturbance through WindAdapter, fused with the state feature.Source: §III · p. 2

A. Wind Disturbance Estimator

The estimator predicts current wind-induced disturbance acceleration from causal proprioceptive history rather than directly evaluating a physical residual online. Its target reflects aerodynamic effects on translational dynamics, and the selected TCN uses a ten-sample history.

A. Wind Disturbance Estimator

  • The disturbance target captures translational acceleration not explained by gravity and nominal collective thrust, including axial-thrust variation and rotor drag from relative airflow.The physical model aggregates parallel rotor axes at the center of mass and omits aerodynamic moments.Source: A. Wind Disturbance Estimator · pp. 2–3
  • Because directly evaluating the physical residual online is sensitive to sensor noise and model mismatch, WAND infers current disturbance acceleration from causal onboard-measurement history.The estimator uses normalized collective-thrust command, body-frame IMU acceleration, and estimated linear velocity among its inputs.Source: A. Wind Disturbance Estimator · p. 3
  • A causal TCN maps a temporal input window to current disturbance acceleration using three residual causal blocks with kernel size 3 and dilations 1, 2, and 4.Left padding ensures the output at time t does not access future measurements.Source: A. Wind Disturbance Estimator · p. 3
  • The estimator uses a ten-sample history, the shortest tested window whose validation RMSE is within 1% of the minimum.The selection came from a validation sweep over window lengths from 4 to 32 samples; the estimate covers disturbance at the vehicle’s current location.Source: A. Wind Disturbance Estimator · p. 3

B. Reinforcement Learning Policy

WAND treats wind-aware navigation as partially observable, augmenting navigation observations with a TCN estimate of disturbance from recent proprioceptive history. The estimate conditions the policy, while the reward links estimated wind direction and magnitude to local obstacle clearance.

B. Reinforcement Learning Policy

  • The policy augments its observation with a TCN disturbance estimate summarizing recent proprioceptive history to address unobserved instantaneous wind.WAND formulates navigation as a partially observable Markov decision process.Source: B. Reinforcement Learning Policy · p. 3
  • A CNN encodes heading-frame 3D LiDAR ray ranges, which are fused with proprioceptive inputs to form the navigation feature.Inputs include goal-relative displacement and vehicle velocity.Source: B. Reinforcement Learning Policy · p. 4
  • The WindAdapter encodes the disturbance estimate into a feature added to the navigation feature, and zero initialization starts its residual at zero.At initialization, the fused feature equals the navigation feature, allowing the wind-conditioned residual to be introduced progressively during joint training.Source: B. Reinforcement Learning Policy · p. 4
  • PPO trains a policy that outputs desired acceleration, bounded before feedforward compensation and transformed into a world-frame setpoint for PX4.The actor's desired acceleration is defined in the heading frame.Source: B. Reinforcement Learning Policy · p. 4
  • The reward combines goal progress, obstacle clearance, height control, wind-aware clearance, and a terminal term; rwind penalizes insufficient clearance along the estimated disturbance direction.For smaller estimated disturbances, rwind is zero; as disturbance increases, its clearance penalty encourages a larger local safety margin.Source: B. Reinforcement Learning Policy · p. 4

C. Simulation and Training Setup

The simulation setup uses randomized analytical wind profiles and dense obstacle environments for training and evaluation. A Gazebo-trained TCN estimator transfers to Isaac Sim without retraining, while the navigation policy is trained jointly with its WindAdapter.

Wind-Profile Generation

  • Four analytical wind profiles vary spatially uniformly; all but constant wind vary over time, and the profiles do not model obstacle-induced wakes or recirculation.The profiles include constant, turbulent, gust, and mixed wind; gusts are finite-duration beta-shaped pulses.Source: C. Simulation and Training Setup · pp. 4–5

Simulation Environments

  • Training uses 1,200 parallel Isaac Sim environments with 500 obstacles each, while Gazebo uses a 20 m × 20 m map with 90 static obstacles.Wind profiles are randomized between episodes; training uses Vmax = 15 m/s, n = 8, and λ = 0.2 s−1.Source: C. Simulation and Training Setup · p. 5

Estimator Training and Transfer

  • The TCN is trained on 90,000 labeled Gazebo samples from constant, turbulent, and gust winds, with a separate mixed-wind dataset held out for evaluation.Episodes are split into disjoint training, validation, and test sets before temporal-window extraction; the estimator transfers to Isaac Sim without retraining and remains frozen during PPO optimization.Source: C. Simulation and Training Setup · p. 5

Policy Training

  • Fig. 4 plots reach-goal rates during PPO training for a navigation-only comparison and WAND ablations, showing five-seed means with 95% confidence intervals.The figure caption describes the comparison and uncertainty display, but does not provide curve values.Source: Figure 4 · p. 5
  • The navigation encoder, WindAdapter, actor, and critic are jointly optimized with PPO for 400 million environment steps.Training takes approximately 3 hours on an NVIDIA RTX 4090 GPU.Source: C. Simulation and Training Setup · p. 5

A. RL Training Results

The ablations indicate that WAND’s late-stage reach-goal improvement depends on combining zero-initialized WindAdapter conditioning with the wind-aware reward. The compared wind-disturbed configurations share environment randomization and PPO settings.

A. RL Training Results

  • The six configurations isolate training with wind, TCN-based feedforward compensation, WindAdapter initialization, and reward choices.WAND combines zero initialization with rwind; the other variants separate these design choices.Source: A. RL Training Results · p. 5
  • All wind-disturbed configurations use the same environment randomization and PPO settings.Source: A. RL Training Results · p. 6
  • Combining zero initialization and rwind gives WAND a clear late-stage reach-goal-rate increase over Base+Comp.Neither zero-initialized WindAdapter without rwind nor rwind with random initialization provides a marked improvement over Base+Comp.Source: A. RL Training Results · p. 6

B. Simulation Results

In simulation, TCN has the lowest disturbance-estimation RMSE across four wind conditions, while WAND improves navigation success over feedforward compensation across the wind-disturbed settings. Opposite-crosswind tests also show trajectory adaptation and higher reported success than Base+Comp.

B. Simulation Results

  • TCN achieves the lowest disturbance-acceleration RMSE under constant, turbulent, gust, and held-out mixed wind.The wind conditions use corresponding episode-level test sets, with mixed wind evaluated on a separately generated held-out composite set.Source: B. Simulation Results · p. 6
  • 8.3 percentage points: WAND’s average observed success-rate improvement over Base+Comp across 12 wind-disturbed settings.WAND has the highest or tied-highest observed success rate in all 13 settings.Source: B. Simulation Results · p. 6, Figure 6 · p. 7
  • Under opposite 7.5 m/s crosswinds, WAND trajectories shift to the opposite side of the central obstacle cluster when wind direction reverses.The fixed-scene comparison overlays 20 trials per method; red crosses and yellow squares mark collisions and timeouts.Source: B. Simulation Results · p. 6, Figure 6 · p. 7
  • WAND succeeds in all trials under both crosswind directions, while Base+Comp records success rates of 60% and 0%.Under +Wx, WAND increases the mean minimum clearance of successful trials by 0.290 m; S/C/T denotes success, collision, and timeout rates.Source: B. Simulation Results · pp. 6–7, Table II · p. 6

C. Real-World Experiments

Indoor fan-induced tests provide real-world evidence of WAND navigation, with 18 successes in 20 trials. Deployment runs onboard at 30 Hz; the policy transfers directly from simulation, while the disturbance estimator is fine-tuned on real-flight data.

C. Real-World Experiments

  • Both onboard modules run at 30 Hz with an average combined inference time of 11.76 ms.The TCN estimator and PPO navigation module contain 196,547 and 146,151 parameters, respectively, for 342,698 total.Source: C. Real-World Experiments · p. 7
  • The navigation policy transfers directly from simulation without further training, while the Gazebo-trained disturbance estimator is fine-tuned on about 10 minutes of real-flight data.On held-out real-flight sequences, the fine-tuned TCN achieves RMSEs of 0.2179 and 0.2442 m/s2 in Scenes 1 and 2, respectively.Source: C. Real-World Experiments · p. 7
  • 18 of 20 trials: WAND succeeds in indoor fan-induced airflow, versus 13 for Base+Comp, 10 for EGO-Planner, and 11 for NavRL.The tests cover two scenes, with locally measured airflow speeds up to approximately 7.2 m/s.Source: C. Real-World Experiments · p. 7, Table III · p. 7

V. CONCLUSIONS

WAND combines a navigation policy, wind-disturbance estimator, and WindAdapter to adapt quadrotor motion under time-varying aerodynamic disturbances. Simulation and controlled indoor fan tests demonstrated improved navigation robustness, while spatially varying wind evaluation remains future work.

V. CONCLUSIONS

  • WAND combines a navigation policy, wind-disturbance estimator, and WindAdapter for quadrotor navigation under time-varying aerodynamic disturbances.The framework targets cluttered environments.Source: §V · p. 7
  • Experiments in simulation and controlled indoor fan tests demonstrated improved navigation robustness.Source: §V · p. 7
  • Systematic evaluation under spatially varying wind fields remains future work.Source: §V · p. 7
Loading 2610.11809v1…