arXiv subjects

Mostafa Eslami

Publications and source records attributed to Mostafa Eslami.

6 recordsLinked to original sources

arXiv2025

Tensor Invariant Data-Assisted Control and Dynamic Decomposition of Multibody Systems

The control of robotic systems in complex, shared collaborative workspaces presents significant challenges in achieving robust performance and safety when learning from experienced or simulated data is employed in the pipeline. A primary bottleneck is the reliance on coordinate-dependent models, which leads to profound data inefficiency by failing to generalize physical interactions across different frames of reference. This forces learning algorithms to rediscover fundamental physical principles in every new orientation, artificially inflating the complexity of the learning task. This paper introduces a novel framework that synergizes a coordinate-free, unreduced multibody dynamics and kinematics model based on tensor mechanics with a Data-Assisted Control (DAC) architecture. A non-recursive, closed-form Newton-Euler model in an augmented matrix form is derived that is optimized for tensor-based control design. This structure enables a principled decomposition of the system into a structurally certain, physically grounded part and an uncertain, empirical, and interaction-focused part, mediated by a virtual port variable. Then, a complete, end-to-end tensor-invariant pipeline for modeling, control, and learning is proposed. The coordinate-free control laws for the structurally certain part provide a stable and abstract command interface, proven via Lyapunov analysis. Eventually, the model and closed-loop system are validated through simulations. This work provides a naturally ideal input for data-efficient, frame-invariant learning algorithms, such as equivariant learning, designed to learn the uncertain interaction. The synergy directly addresses the data-inefficiency problem, increases explainability and interpretability, and paves the way for more robust and generalizable robotic control in interactive environments.

cs.RO↗

arXiv2025

Learning-Based Data-Assisted Port-Hamiltonian Control for Free-Floating Space Manipulators

A generic data-assisted control architecture within the port-Hamiltonian framework is proposed, introducing a physically meaningful observable that links conservative dynamics to all actuation, dissipation, and disturbance channels. A robust, model-based controller combined with a high-gain decentralized integrator establishes large robustness margins and strict time-scale separation, ensuring that subsequent learning cannot destabilize the primary dynamics. Learning, selected for its generalizability, is then applied to capture complex, unmodeled effects, despite inherent delay and transient error during adaptation. Formal Lyapunov analysis with explicit stability bounds guarantees convergence under bounded learning errors. The structured design confines learning to the simplest part of the dynamics, enhancing data efficiency while preserving physical interpretability. The approach is generic, with a free-floating space manipulator orientation control task, including integrated null-space collision avoidance, serving as a case study to demonstrate robust tracking performance and applicability to broader robotic domains.

eess.SY↗

arXiv2025

On the Generalization of Data-Assisted Control in port-Hamiltonian Systems (DAC-pH)

This paper introduces a hypothetical hybrid control framework for port-Hamiltonian (p$\mathcal{H}$) systems, employing a dynamic decomposition based on Data-Assisted Control (DAC). The system's evolution is split into two parts with fixed topology: Right-Hand Side (RHS)- an intrinsic Hamiltonian flow handling worst-case parametric uncertainties, and Left-Hand Side (LHS)- a dissipative/input flow addressing both structural and parametric uncertainties. A virtual port variable $\Pi$ serves as the interface between these two components. A nonlinear controller manages the intrinsic Hamiltonian flow, determining a desired port control value $\Pi_c$. Concurrently, Reinforcement Learning (RL) is applied to the dissipative/input flow to learn an agent for providing optimal policy in mapping $\Pi_c$ to the actual system input. This hybrid approach effectively manages RHS uncertainties while preserving the system's inherent structure. Key advantages include adjustable performance via LHS controller parameters, enhanced AI explainability and interpretability through the port variable $\Pi$, the ability to guarantee safety and state attainability with hard/soft constraints, reduced complexity in learning hypothesis classes compared to end-to-end solutions, and improved state/parameter estimation using LHS prior knowledge and system Hamiltonian to address partial observability. The paper details the p$\mathcal{H}$ formulation, derives the decomposition, and presents the modular controller architecture. Beyond design, crucial aspects of stability and robustness analysis and synthesis are investigated, paving the way for deeper theoretical investigations. An application example, a pendulum with nonlinear dynamics, is simulated to demonstrate the approach's empirical and phenomenological benefits for future research.

eess.SY↗

arXiv2024

Particle Filter Optimization: A Bayesian Approach for Global Stochastic Optimization

This paper proposes a novel global optimization algorithm, Particle Filter-Based Optimization (PFO), designed for a class of stochastic optimization problems in which the objective function lacks an analytical form and is subject to noisy evaluations. PFO utilizes the Bayesian inference framework of Particle Filters (PF) by reformulating the optimization task as a state estimation problem. In this context, evaluations of the objective function are interpreted as measurements, and a customized transition model based on covariance ellipsoids is introduced to guide particle propagation. This model serves as a surrogate for classical acquisition functions, equipping the PF framework with local search capabilities and supporting efficient exploration of the global optimum. To mitigate the adverse effects of measurement noise, the Unscented Transform (UT) is employed to approximate the underlying mean of the objective function, enhancing the accuracy of particle updates. The algorithm offers notable improvements over existing stochastic optimization algorithms for black-box multi-modal objective functions. First, PFO provides a fully probabilistic definition of particle weights, enhancing adaptability and robustness. Second, PFO integrates exploration and exploitation within a unified Bayesian framework, ensuring a non-zero probability of sampling from unexplored regions throughout the optimization process. This approach contrasts with traditional particle filter methods that are primarily used for state estimation, and heuristic optimization algorithms that lack theoretical guarantees. The novelty of PFO lies in its unique integration of particle filtering with a dynamic search space prediction, offering a theoretically grounded alternative to acquisition functions in Bayesian Optimization (BO).

math.OC↗

arXiv2023

Sequential Data-Assisted Control in Flight

Flight dynamics involve uncertainties in parameters, aerodynamic derivatives, and engine thrust. These uncertainties can be categorized into three types: known-predictable, known-unpredictable, and unknown. While advanced control systems typically rely on high-fidelity dynamical models in dealing with known-predictable uncertainties, simplified approaches are used for the second and third categories to manage the complexities involved in synthesis and implementation. In this paper, the focus is on accurately modeling the internal dynamics, which primarily deal with parametric uncertainties. Real-time data is employed to identify uncertainties in the remaining external dynamics, including both known-unpredictable and unknown aspects. To address these uncertainties and maintain optimal performance, stability, and robustness, the authors propose a framework known as Sequential Data-Assisted Control (SDAC). This framework involves using a model-based nonlinear controller for the internal dynamics to provide the desired momentum to a data-based controller responsible for the external dynamics. By the momentum through the internal dynamics and leveraging the Koopman operator, the linear evolution of momentum is derived. This information is then utilized by the data-based controller to assign appropriate control inputs. The proposed approach establishes a novel foundation for a comprehensive analysis of maneuverability, stabilizability, and controllability. To evaluate the performance of SDAC, closed-loop simulations are conducted using the NASA Generic Transport Model (GTM). The data-based controller employs Linear Quadratic Regulator (LQR), while the model-based controller uses robust sliding mode control. A comparison is made with a pure robust nonlinear controller, demonstrating significant performance improvements, particularly in cases involving known-unpredictable uncertainties.

eess.SY↗

arXiv2023

Data-Assisted Control -- A Framework Development by Exploiting NASA GTM Platform

Today's focus on expanding the capabilities of control systems, resulting from the abundance of data and computational resources, requires data-based alternatives over model-based ones. These alternatives may become the sole tool for analysis and synthesis. Nevertheless, mathematical models are available to some extent, especially for air and space vehicles. Hypothetically, data assistance would be the approach to meet the requirements in collaboration with the model. In this paper, a framework of Data-Assisted Control (DAC) for aerospace vehicles is proposed. NASA Generic Transport Model (GTM) is the platform for the study and the data supports the model-based controller in extending performance over a damage event. The framework requires real-time decisions to override the control law with the information obtained from the data, while the model-based controller does not show regular performance. The closed-loop system is shown to be stable in the transition phase between the data and the model. The fixed dynamic parameters are estimated using the Dual Unscented Kalman Filter (DUKF) and the evolution of the generalized force moments is estimated using the Koopman estimator. Simulations have shown that the purely model-based robust control leads to degradation of the closed-loop performance in case of damage, suggesting the need for data assistance.

eess.SY↗