PaperlayerView exact revision on arXiv

Source-linked AI summary

Prediction-Informed Landing Window Planning for UAV Landing on USV in High Sea States

Kaizhen Li, Xi Zhou, Dan Zhang, Xiaowei Li

arXiv:2610.13190v1cs.ROcs.LG

TL;DR

High-sea-state UAV landing on a USV is difficult because rapidly changing deck motion can make a safe touchdown condition unfavorable before contact, while existing hierarchical methods infer future motion implicitly from short histories. This paper adds pretrained attitude prediction to hierarchical landing-window planning; real-sim results report improved awareness of future windows under rapidly varying conditions.

  • Problem

    Existing hierarchical landing methods infer future platform motion implicitly from short observation histories, limiting their awareness of landing windows in rapidly changing high-sea-state conditions.

    Source: §II · p. 2

  • Method

    A pretrained temporal model forecasts USV attitude for high-level landing-window decisions, while conventional low-level control stabilizes the UAV and tracks motion commands.

    Source: §I · p. 2

  • Results

    Real-sim results show that explicit attitude prediction improves policy awareness of future landing windows and is more suitable for rapidly varying sea conditions.

    Source: §V · p. 4

  • Takeaways & Limitations

    The policy can continuously adjust landing timing with descent-rate changes, hovering, or opposite-direction velocity commands, without manually designed state-switching rules.

    Source: §V · p. 4

    Because high-sea-state data are limited, prediction and planning consider only USV pitch and roll; physical UAV-USV validation remains future work.

    Source: §V · p. 4

Abstract

from arXiv · show

Autonomous landing of unmanned aerial vehicle (UAV) on unmanned surface vehicle (USV) is difficult in high sea states. Existing hierarchical landing methods often infer future USV motion only implicitly from short observation histories, which limits their ability to adapt to severe and rapidly changing sea conditions. This paper presents a prediction-informed hierarchical control framework. A pretrained model is developed to explicitly predict the USV attitude during high-level landing decisions. The prediction helps identify suitable landing windows and choose the right touchdown time, while a conventional low-level flight controller maintains UAV stability and tracks motion commands. To evaluate the framework under realistic disturbances, a real-sim validation setup is constructed in which USV-motion data collected in real sea conditions are transferred to a MuJoCo USV. The experiments demonstrate the feasibility of using explicit attitude prediction to improve landing-window selection and support timing-aware UAV landing planning in rapidly varying marine environments.

I. INTRODUCTION

High sea states make UAV-on-USV landing a timing problem: deck motion can turn a currently safe touchdown unfavorable before contact. The paper addresses this with explicit attitude prediction for landing-window decisions and a real-to-simulation evaluation setup.

I. INTRODUCTION

  • Reliable autonomous landing supports sustained UAV-USV operation by enabling recovery, recharging, and redeployment without continuous human intervention.The UAV provides aerial observation and coverage, while the USV provides endurance and access to the water surface.Source: §I · p. 1
  • A touchdown condition that is safe now may become unfavorable before contact as waves change the USV’s heave, roll, and pitch.Wind, sensing uncertainty, and control delay further reduce the available touchdown margin.Source: §I · p. 1
  • The framework uses a pretrained temporal model to forecast USV attitude for touchdown-window assessment while a conventional controller stabilizes and tracks the UAV.Predicted attitude guides the high-level landing decision.Source: §I · p. 2
  • Real-sea USV-motion data are replayed on a MuJoCo platform to evaluate the framework under realistic disturbance patterns.The setup transfers platform-motion data collected in real sea conditions to the simulation.Source: Figure 1 · p. 1, §I · p. 2

II. RELATED WORK

Prior hierarchical landing methods retain conventional stabilization but often infer future platform motion implicitly from short observation histories. This work instead exposes predicted attitude to high-level landing-window planning.

II. RELATED WORK

  • Earlier learning-based work spans visual end-to-end landing, moving-platform landing, and hierarchical policies that pair learned decisions with conventional flight control.WaveLander uses a compact platform-relative observation and outputs a vertical-velocity reference while the low-level controller handles stabilization and tracking [8].Source: §II · p. 2
  • Hierarchical methods can slow, hold, or retreat under unfavorable platform conditions, but their future-motion prediction remains implicit and relies on short-term evidence.Under rapidly changing high-sea-state conditions, an unfavorable touchdown may be recognized only after a safe interval starts closing.Source: §II · p. 2
  • The proposed approach treats attitude prediction as a high-level information source for assessing touchdown windows, separate from synchronization or tracking control.The conventional low-level flight-control stack is retained, so the learned component does not directly control UAV actuators.Source: §II · p. 2

III. METHODOLOGY

The system combines real-sea USV-motion data, a pretrained short-term attitude predictor, and a hierarchical controller. PPO chooses a vertical-velocity reference, while low-level control and reward design support safe landing timing.

A. System Overview

  • Real-sea USV motion trains a model to predict short-term attitude changes, and the motion data drive the dynamic platform in MuJoCo.The architecture combines the real-data-trained prediction module with a simulated landing environment.Source: A. System Overview · p. 2
  • PPO receives current deck tilt and tilt rate, UAV relative height and vertical velocity, and predicted attitude as its observation.The predicted attitude is combined with the current landing state.Source: A. System Overview · p. 2
  • PPO outputs a continuous vertical-velocity command; the SE(3) controller allocates motor thrusts, while reward terms cover alignment, descent progress, touchdown safety, and platform attitude.The reward design guides selection of an appropriate landing time.Source: A. System Overview · p. 3

B. Pretrained Module

An encoder-decoder LSTM forecasts short-term USV motion before RL optimization. The prediction-informed model combines forecast deck motion with current landing state for PPO, while an SE(3) controller handles low-level flight control.

B. Pretrained Module

  • The encoder-decoder LSTM uses USV motion sampled at 0.1 s, with 20 historical samples predicting the next 10 samples for a 1.0 s forecast.During online inference, the forecast is updated every 0.1 s and sent to the RL environment.Source: B. Pretrained Module · p. 3
  • The prediction module is trained before RL optimization and uses an independently normalized input window, MSE loss, and Adam with weight decay, scheduling, and early stopping.MAE is used for monitoring.Source: B. Pretrained Module · p. 3

C. Prediction-Informed Model

  • PPO’s observation combines current relative height, vertical velocity, platform tilt, and tilt rate with ten predicted roll samples.The predicted sequence informs landing-window awareness and reward evaluation but does not directly replace the policy action.Source: C. Prediction-Informed Model · p. 3
  • PPO outputs only a vertical-velocity reference; the SE(3) controller computes desired acceleration, thrust direction, attitude, and moments, then allocates the wrench to four motors.The low-level controller maintains stabilization and tracks the high-level reference.Source: C. Prediction-Informed Model · p. 3

D. Reward Design

The reward is an additive function combining descent progress, touchdown, forecast, action-rate, and safety terms. These components shape descent toward contact while accounting for predicted platform motion and command changes.

D. Reward Design

  • The total reward adds progress, touchdown, forecast, action-rate, and safety components.The reward is defined as the sum of these five terms.Source: D. Reward Design · p. 3
  • The progress term tracks a height-dependent target descent velocity, while the touchdown term activates only within specified height and horizontal-error thresholds.The target descent velocity is faster at high altitude and approaches zero near contact; touchdown reward activates when height is at most 0.50 m and horizontal error at most 0.25 m.Source: D. Reward Design · p. 3
  • The forecast term uses predicted-attitude stability and peak-tilt excess, and the action-rate term penalizes abrupt changes in vertical-velocity commands.Source: D. Reward Design · p. 3
  • The safety term penalizes excessive tilt and descent speed; episode termination awards +5000 for successful landing and −2000 otherwise.The stated thresholds are tilt above 15 degrees and descent speed below −1.0 m/s.Source: D. Reward Design · p. 3

E. Training Setup

The predictor is trained on high-sea-state USV motion using temporally ordered input-target windows. The landing policy is trained in MuJoCo with recorded or generated motion and PPO.

E. Training Setup

  • Training windows use 20 historical samples to predict the following 10 samples, with 0.1 s sampling yielding a 2.0 s history and 1.0 s forecast.The dataset is split 7:3 between training and test subsets.Source: E. Training Setup · p. 3
  • The MuJoCo policy-training episode starts at 2.0 m altitude with randomized horizontal position, observes relative motion and deck state plus a 10-step roll forecast, and outputs vertical-velocity commands.The randomized horizontal position is within [−2, 2] m, and commands range from −0.8 to 0.8 m/s.Source: E. Training Setup · p. 3
  • PPO optimizes the policy using two-layer actor and critic networks with 128 units per layer and the reported rollout, batch, and update settings.The learning rate is 3 × 10−4, rollout length 2048 steps, batch size 64, and each rollout is updated for 10 epochs.Source: E. Training Setup · p. 3

A. Pretrained Module Behavior Analysis

The pretrained predictor processes roll and pitch histories and produces a 10-step sequence. Its test-set forecasts track measured attitude over a 1.0 s horizon, and LSTM has the lowest reported errors among the compared architectures.

A. Pretrained Module Behavior Analysis

B. Policy Behavior Analysis

The policy uses current landing state and predicted deck attitude to adjust vertical velocity. It changes descent timing ahead of unfavorable tilt and approaches contact during a more favorable interval.

B. Policy Behavior Analysis

  • PPO uses current landing state and predicted deck attitude to descend when the platform is favorable and slow, hover, or briefly ascend when a large attitude change is predicted.Source: B. Policy Behavior Analysis · p. 4
  • The policy changes from descent to upward command at about 4 s before deck tilt rises rapidly around 5 s, avoiding entry into that unfavorable window.Around 17–18 s, it maintains near-zero vertical velocity when tilt is small before approaching the contact threshold.Source: B. Policy Behavior Analysis · p. 4
  • Across the motion segments, the UAV vertical-velocity command leads USV attitude changes, prompting earlier slowdown, hovering, or climbing before rapid deck-angle increases.Fig. 5 segments show height, PPO vertical-velocity command and actual velocity, and USV deck tilt; dashed lines mark contact or safety thresholds.Source: B. Policy Behavior Analysis · p. 4

V. CONCLUSION

The framework combines explicit USV attitude prediction with hierarchical control to support landing-window planning in high sea states. Real-sim results indicate that prediction improves future-window awareness, while the current framework considers only USV pitch and roll.

V. CONCLUSION

  • The framework forecasts USV roll and pitch for PPO landing decisions, while a low-level SE(3) controller tracks vertical-velocity references and stabilizes the UAV.Source: §V · p. 4
  • Real-sim results show that explicit attitude prediction improves future landing-window awareness and supports continuous landing-timing adjustments.Fig. 5 tracks UAV height, commanded and actual vertical velocity, and deck tilt against contact or safety thresholds.Source: §V · p. 4, Figure 5 · p. 5
  • Limited high-sea-state data restrict prediction and planning to USV pitch and roll; heave, displacement, and physical-platform validation remain future work.Source: §V · p. 4
Loading 2610.13190v1…