Learning Long-Horizon Predictions for Quadrotor Dynamics

TL;DR

This work tackles the challenge of long-horizon quadrotor dynamics by analyzing how history, multi-step loss, and sequential architectures affect compounding prediction errors. It introduces a decoupled dynamics learning framework with separate Velocity and Attitude predictors and demonstrates that a Temporal Convolutional Network (TCN) backbone, combined with history and multi-step supervision, yields the most accurate and stable long-horizon forecasts. Through extensive experiments on real-world datasets PI-TCN and NeuroBEM, the authors show substantial improvements in both velocity and attitude prediction over short-horizon baselines and other predictor types, validating the approach for planning and control. The findings offer practical guidelines for designing data-driven dynamics models that maintain accuracy over long horizons, with potential to improve model-based planning and control in UAV applications.

Abstract

Accurate modeling of system dynamics is crucial for achieving high-performance planning and control of robotic systems. Although existing data-driven approaches represent a promising approach for modeling dynamics, their accuracy is limited to a short prediction horizon, overlooking the impact of compounding prediction errors over longer prediction horizons. Strategies to mitigate these cumulative errors remain underexplored. To bridge this gap, in this paper, we study the key design choices for efficiently learning long-horizon prediction dynamics for quadrotors. Specifically, we analyze the impact of multiple architectures, historical data, and multi-step loss formulation. We show that sequential modeling techniques showcase their advantage in minimizing compounding errors compared to other types of solutions. Furthermore, we propose a novel decoupled dynamics learning approach, which further simplifies the learning process while also enhancing the approach modularity. Extensive experiments and ablation studies on real-world quadrotor data demonstrate the versatility and precision of the proposed approach. Our outcomes offer several insights and methodologies for enhancing long-term predictive accuracy of learned quadrotor dynamics for planning and control.

Figures (5)

• Figure 1: Illustration of how errors in dynamical models accumulate over time. Each subsequent prediction integrates the errors from all previous steps, leading to increased cumulative error and reduced accuracy in long-term predictions.
• Figure 2: Full State Predictor predicts complete state vectors, Multi Head Predictor uses separate heads for velocity and attitude prediction, while our proposed Decoupled Predictor decomposes the problem into velocity and attitude prediction modules, introducing a novel approach for enhanced modularity and long-horizon prediction.
• Figure 3: Selecting model parameters ensuring real-time performance on embedded systems. The parameter bound is determined by selecting the model with the lowest parameter count capable of real-time performance. LSTM achieves real-time predictions with up to $5.2$ million parameters. We freeze this bound across all models to have a fair comparison.
• Figure 4: Mean Squared Error per-step among baseline neural networks models on unseen test trajectories. The MLP model demonstrates higher error variance across diverse unseen trajectories compared to other sequential models, attributed to the absence of architectural priors for capturing spatio-temporal dependencies across datasets.
• Figure 5: Mean and standard deviation of the best-performing predictors across $10$ random trials, evaluating velocity and attitude error. A, B, and C correspond to Full-State, Multi-Head, and Decoupled, respectively.