AERMANI-Diffusion: Regime-Conditioned Diffusion for Dynamics Learning in Aerial Manipulators

Aerial manipulators undergo rapid, configuration-dependent changes in inertial coupling forces and aerodynamic forces, making accurate dynamics modeling a core challenge for reliable control. Analytical models lose fidelity under these nonlinear and nonstationary effects, while standard data-driven methods such as deep neural networks and Gaussian processes cannot represent the diverse residual behaviors that arise across different operating conditions. We propose a regime-conditioned diffusion framework that models the full distribution of residual forces using a conditional diffusion process and a lightweight temporal encoder. The encoder extracts a compact summary of recent motion and configuration, enabling consistent residual predictions even through abrupt transitions or unseen payloads. When combined with an adaptive controller, the framework enables dynamics uncertainty compensation and yields markedly improved tracking accuracy in real-world tests.

💡 Research Summary

This paper introduces AERMANI-Diffusion, a novel framework for learning and compensating for the complex, regime-dependent dynamics of aerial manipulators (quadrotors equipped with robotic arms). The core challenge addressed is the rapid, configuration-dependent variation in inertial coupling and aerodynamic forces, which renders traditional analytical models inaccurate and poses difficulties for standard data-driven approaches like Deep Neural Networks (DNNs) or Gaussian Processes (GPs). These methods typically learn a single, deterministic mapping or assume stationary, smooth dynamics, failing to capture the multimodal distribution of residual forces that arise across different operational conditions such as payload changes, manipulator postures, and flight phases.

The authors propose a two-part solution. First, they formally reframe the dynamics learning problem. Instead of modeling each physical term separately, they consolidate all unmodeled nonlinear, nonstationary, and disturbance effects into a single “residual force” vector (H). Analysis of real flight data, visualized via t-SNE embeddings, reveals that these residual forces cluster into distinct groups based on operating regime, confirming that the mapping from state/input to residual is not unique but depends on latent context.

Second, to model this regime-dependent behavior, the paper presents a regime-conditioned diffusion model. A key innovation is the use of a lightweight Temporal Convolutional Network (TCN) encoder that processes a short history of past states and control inputs. This encoder extracts a compact “regime descriptor” (r_t) that encapsulates contextual information about the current operating condition (e.g., whether a payload is attached, the manipulator’s recent motion). This descriptor conditions a Denoising Diffusion Probabilistic Model (DDPM), which learns the full conditional distribution p(H | state, input, regime). This allows the model to generate consistent residual predictions appropriate for the current regime, preventing it from averaging across incompatible behaviors—a flaw demonstrated in ablation studies.

For closed-loop control, the learned model is integrated with an adaptive controller. The controller uses the diffusion model’s residual estimate (Ĥ) for feedforward compensation. Crucially, it also incorporates an adaptive term that online estimates and counters the bounded residual estimation error (σ), providing robustness against model inaccuracies and disturbances. A theoretical analysis proves that the resulting closed-loop system is Uniformly Ultimately Bounded (UUB).

The framework is validated through real-world experiments on an aerial manipulator platform. Results show that the regime-conditioned diffusion approach, combined with adaptive control, significantly outperforms baseline methods (including standard DNNs and non-conditioned diffusion models) in trajectory tracking accuracy. It demonstrates robust performance across challenging scenarios involving payload pickup/release, fast manipulator motions, and trajectories unseen during training. The work successfully bridges advanced generative modeling with adaptive control theory, offering a principled and effective method for handling the complex, switching dynamics inherent in next-generation robotic systems.