Wake Up to the Past: Using Memory to Model Fluid Wake Effects on Robots

Luca Vendruscolo, Eduardo Sebastián, Amanda Prorok, Ajay ShankararXiv 2026 · arXiv

TL;DR

When robots fly or swim in close proximity, the fluid wakes they generate create severe aerodynamic or hydrodynamic disturbances for each other. Because these fluid effects take time to travel through space, standard "memory-less" neural networks fail to predict them during agile maneuvers. This paper demonstrates that incorporating temporal history and explicitly modeling transport delays drastically improves disturbance prediction accuracy across both aerial and aquatic environments.

What Problem Does This Paper Address?

Imagine a fleet of quadrotors flying in tight formation or autonomous underwater vehicles docking in a harbor. As they move, they push against the air or water, creating chaotic fluid wakes (like downwash or trailing vortices). A nearby "sufferer" robot entering this wake experiences sudden, destabilizing forces.

To safely compensate for these forces in real-time, robots need fast, data-driven predictors. Traditionally, researchers have used memory-less neural networks that map the current relative positions and velocities of the robots directly to the current force experienced. However, this assumption falls apart during fast, agile maneuvers. Because fluid disturbances travel at a finite speed, the force hitting a sufferer robot right now is actually a function of where the "producer" robot was at some point in the past. Without temporal context, memory-less models suffer from severe ambiguity and perform poorly.

Methods

The authors conduct a comprehensive empirical study evaluating how different temporal mechanisms capture fluid dynamics. They restrict all models to under 100K trainable parameters to ensure they remain computationally lean enough for real-time, on-board deployment.

They test seven data-driven architectures categorized into four groups:

  1. Memory-less Baseline: Agile MLP, which only looks at instantaneous features.
  2. Explicit History Models: History MLP (uses a flattened sliding window of past states) and Delay Embedding (uses a learned Gaussian attention mechanism to select a specific past snapshot).
  3. Recurrent Models: Gated Recurrent Unit (GRU), Temporal Convolutional Network (TCN), Mamba (a selective state-space model), and Echo State Networks (RC-ESN).
  4. Attention-based Models: Cross-Attention, which allows the sufferer to dynamically query relevant moments from the producer's past trajectory.

Datasets

The architectures were evaluated across four diverse simulated/experimental domains and one real-world hardware setup:

  • FlareDW CFD: A high-fidelity computational fluid dynamics dataset capturing aggressive, non-stationary downwash interactions between two quadrotors at speeds up to 4.0 m/s.
  • Ship Encounter: A numerical simulation modeling the sway forces and yaw moments experienced by ships during close-proximity overtaking maneuvers in restricted channels.
  • Fish Schooling: Real experimental data from a robotic setup consisting of an actively pitching airfoil shedding vortices onto a passive downstream flag.
  • 2D Downwash Simulator: A custom Eulerian-grid fluid simulation used to test closed-loop trajectory tracking and out-of-distribution generalization.
  • Real Drone Gantry (Hardware Validation): A custom-built planar rectilinear gantry featuring two spinning monocopters interacting under time-varying thrust conditions.

Key Insights and Ideas

  • No Single Architecture Wins Everywhere: The optimal temporal mechanism depends heavily on the physical properties of the medium and the motion patterns.
  • The Danger of Periodic Aliasing: In the Fish Schooling domain, the disturbances are highly sinusoidal. Explicit delay models (like Delay Embedding) often fail here because they get trapped in local minima ("alias peaks") where shifted copies of the periodic signal correlate well but don't represent the true physical delay. Recurrent models like GRUs handle this periodic phase ambiguity much better.
  • Stationary vs. Non-Stationary Delays: When the distance between robots is roughly constant (like quadrotors in vertical formation), explicit delay prediction via Cross-Attention shines. However, when robots approach and recede (like the Ship Encounter), the transport delay continuously changes. In these non-stationary scenarios, a simple sliding window (History MLP) performs best because it easily maps the encounter phase.

Key Results

Across all domains, models incorporating temporal context substantially outperformed the memory-less baseline.

Fig. 1: Comparing models with and without memory: we study the influence of modeling temporal context in the prediction of fluid wake effects in four domains, including aerial and aquatic scenarios. In dynamic regimes, data-driven predictors must support history and learn the transport delay from the medium. Illustrations of the experimental domains in the third and fourth column are extracted from [1] and [2].
  • Quadrotor CFD: Cross-Attention reduced the downwash force prediction Root Mean Squared Error (RMSE) from 1.78N (Memory-less) down to 1.15N.
  • Fish Schooling: While the memory-less MLP achieved a low absolute error, its coefficient of determination ($R^2$) was only 0.57 due to phase ambiguity. The GRU resolved this, achieving an $R^2$ of 0.96.
  • Real-World Hardware: Open-loop testing on the physical monocopter gantry confirmed that the History MLP and TCN provided the cleanest, lowest-error force predictions, successfully accounting for both fluid transport delays and the mechanical spool-up lag of the motors.

Limitations and Future Work

The primary limitation identified is the lack of a single "silver bullet" architecture that dominates every domain. While the paper proves that an accurate predictor must represent history and model transport delays, combining these two mechanisms into a unified, universally robust module remains an open challenge.

Additionally, the current framework is strictly evaluated on pairwise interactions (one producer, one sufferer). The authors note that extending this to multi-agent swarms—where emergent fluid interactions are far more complex—is a critical next step for future research.

Summary updated

Abstract

Autonomous aerial and aquatic robots that attain mobility by perturbing their medium, such as multicopters and torpedoes, produce wake effects that act as disturbances for adjacent robots. Wake effects are hard to model and predict due to the chaotic spatio-temporal dynamics of the fluid, entangled with the physical geometry of the robots and their complex motion patterns. Data-driven approaches using neural networks typically learn a memory-less function that maps the current states of the two robots to a force observed by the "sufferer" robot. Such models often perform poorly in agile scenarios: since the wake effect has a finite propagation time, the disturbance observed by a sufferer robot is some function of relative states in the past. In this work, we present an empirical study of the properties a wake-effect predictor must satisfy to accurately model the interactions between two robots mediated by a fluid. We explore seven data-driven models designed to capture the spatio-temporal evolution of fluid wake effects in four different media. This allows us to introspect the models and analyze the reasons why certain features enable improved accuracy in prediction across predictors and fluids. As experimental validation, we develop a planar rectilinear gantry for two spinning monocopters to test in real-world data with feedback control. The conclusion is that support of history of previous states as input and transport delay prediction substantially helps to learn an accurate wake-effect predictor.
BibTeX
@misc{vendruscolo2026wake,
  author = {Luca Vendruscolo and Eduardo Sebastián and Amanda Prorok and Ajay Shankar},
  title = {{Wake Up to the Past: Using Memory to Model Fluid Wake Effects on Robots}},
  year = {2026},
  eprint = {2603.22472},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.22472},
}
Figure