Technical Summary: Learning Cross-Coupled and Regime Dependent Dynamics for Aerial Manipulation

Problem Statement

Aerial manipulators, which combine multirotor vehicles with articulated robotic arms, face significant challenges in dynamic modeling due to strong quadrotor-manipulator coupling, delayed aerodynamic interactions, and regime-dependent variations. These variations arise from payload changes (e.g., pickup and release) and manipulator reconfiguration, leading to residual dynamics that are simultaneously cross-coupled, history-dependent, and nonstationary.

Existing approaches struggle in this domain:

• Analytical models often fail to capture complex inertial and aerodynamic couplings.
• Purely offline learned models (e.g., standard DNNs, GPs) degrade during deployment because they cannot adapt to distribution shifts caused by changing operating regimes.
• Online adaptation methods often rely on gradient descent (requiring careful tuning and suffering from poor sample efficiency) or full network retraining (introducing computational overhead). Furthermore, many existing methods do not simultaneously address strong cross-variable coupling and abrupt, regime-dependent nonstationarity.

The core challenge is to model residual dynamics that capture tight coupling between state variables and temporal dependencies while enabling rapid, stable online adaptation to unseen operating conditions without compromising real-time control performance.

Methodology

The authors propose a structured encoder-decoder framework for adaptive residual dynamics learning. The system approximates the residual dynamics $f_\Delta$ as a linear combination of learned latent features: $f_\Delta(H_k) \approx \Theta z_k$.

1. Nonlinear Latent Encoder (Offline Trained)

The encoder $\phi(\cdot)$ processes a history matrix $H_k$ containing recent state-input trajectories. It is designed to capture two specific characteristics identified in the problem analysis:

• Cross-Variable Coupling: Modeled via cross-variable attention blocks that compute dependencies across different state and input channels within temporal segments.
• Temporal Dependencies: Modeled via temporal attention blocks that capture delayed effects (e.g., aerodynamic wake, actuator dynamics) across time steps.
• Architecture: The encoder uses alternating temporal and cross-variable attention layers, followed by attention-based pooling and a projection network to produce a compact latent representation $z_k \in \mathbb{R}^\ell$. This encoder is trained offline and remains fixed during deployment.

2. Linear Latent Decoder (Online Adapted)

The decoder matrix $\Theta$ maps the latent features to the residual target. Crucially, the model is linear-in-parameters with respect to $\Theta$, enabling efficient closed-form Bayesian updates.

• Recursive Update: The decoder parameters are updated recursively using a Kalman filter in parameter space. The parameter evolution is modeled as a random walk, allowing the estimator to track time-varying dynamics.
• Consistency-Driven Covariance Inflation: To handle abrupt regime shifts (e.g., payload release), the system monitors the normalized innovation score (Mahalanobis distance between predicted and observed residuals).
  • An Exponentially Weighted Moving Average (EWMA) of this score ($g_k$) detects persistent mismatches.
  • If a mismatch is detected, an adaptive covariance inflation mechanism scales the prior covariance by a factor $1/\lambda_k$ (where $\lambda_k < 1$). This effectively acts as an adaptive forgetting factor, increasing the Kalman gain to accelerate adaptation without resetting the estimator.

3. Model Predictive Control (MPC) Integration

The learned residual model is embedded into a nonlinear MPC framework. During the prediction horizon, the encoder is frozen, and the residual is treated as a constant additive correction based on the current posterior mean of the decoder. This formulation ensures computational tractability for real-time control while accounting for unmodeled dynamics.

Key Contributions

1. Nonlinear Residual Representation Learning: A history-dependent latent encoder that explicitly models cross-variable coupling and temporal dependencies, addressing the limitations of instantaneous or purely temporal models.
2. Online Regime-Adaptive Residual Estimation: A linear latent decoder structure that enables lightweight, closed-form Bayesian adaptation. The inclusion of consistency-driven covariance inflation allows for rapid, stable adaptation to both transient and slowly varying dynamics changes, specifically tailored for the nonstationary nature of aerial manipulation.

Experimental Results

The framework was validated on a real-world aerial manipulator platform (Tarot-650 quadrotor with a 2R serial-link manipulator) under two challenging scenarios involving payload release and trajectory switching:

• Open-Loop Prediction: The proposed "Full Structured" encoder outperformed baselines (DNN, TCN, and temporal-only encoders) with the lowest RMSE (0.18 vs. 0.21–0.34), demonstrating the necessity of modeling both cross-variable and temporal interactions.
• Decoder Ablation: The full adaptive decoder (with consistency-driven inflation) achieved the lowest RMSE and fastest recovery after an abrupt dynamics shift (payload drop) compared to fixed decoders, standard Bayesian updates, and fixed forgetting factor methods.
• Trajectory Tracking: In closed-loop MPC experiments, the proposed method outperformed nominal models, offline DNNs, diffusion-based models, and other online adaptation baselines (e.g., Active MLP, Proto-MPC). It demonstrated superior disturbance rejection and tracking accuracy under both in-distribution and out-of-distribution payload conditions.

Significance and Claims

The paper claims that accurate modeling of aerial manipulators requires the joint modeling of coupled temporal dynamics and deployment-time nonstationarity. The proposed framework addresses this by decoupling the complex representation learning (handled offline by the nonlinear encoder) from the adaptation mechanism (handled online by the linear decoder).

The authors emphasize that their approach enables:

• Rapid adaptation to abrupt changes (like payload release) without the instability often associated with aggressive gradient-based learning.
• Stable estimation under nominal conditions due to the conservative nature of the Bayesian update when no mismatch is detected.
• Real-time compatibility with MPC, as the online update is computationally efficient ($O(d\ell^2)$) and the encoder inference is performed once per step.

The results highlight that combining structured history encoding with adaptive covariance inflation provides a robust solution for reliable aerial manipulation in varying and unpredictable operating regimes.