neural extended state observer

This work addresses the limited robustness of learning-based control under strong disturbances and out-of-distribution scenarios, where overreliance on learned models often leads to performance degradation. To overcome this, the authors propose Neural-ESO, a dual-path architecture that combines a neural network-based feedforward predictor for rapid disturbance estimation with an extended state observer (ESO) that corrects prediction errors, thereby reducing dependence on the learning component. By innovatively integrating neural networks with ESO and incorporating Lipschitz continuity constraints, the method establishes, for the first time, uniform ultimate boundedness of the closed-loop error dynamics, offering theoretically guaranteed robustness. Evaluated on a quadrotor landing task under strong ground effect, Neural-ESO consistently outperforms state-of-the-art baselines across training, deployment, and transfer phases, achieving a superior balance between accuracy and reliability.

To address the challenge of high-precision control in robotic systems arising from inaccurate dynamic modeling and strong coupling between internal and external disturbances, this paper proposes a learning-enhanced high-order disturbance observer (HODO). The method innovatively integrates Chebyshev orthogonal series expansion with regularized least-squares (RLS) online learning within the HODO framework, enabling provably convergent, high-accuracy real-time estimation of coupled uncertainties and exogenous disturbances. It requires no prior knowledge of disturbance models and achieves both strong robustness and computational efficiency. Simulation results demonstrate rapid convergence of estimation errors and significant improvement in closed-loop trajectory tracking accuracy. This work establishes a novel paradigm for learning-augmented control of nonlinear systems and provides a theoretically verifiable analytical tool grounded in rigorous convergence guarantees.

To address the challenges of unmeasurable states, imprecise models, and absence of ground-truth labels in nonlinear systems, this paper proposes a physics-informed adaptive sliding-mode neural observer. The method integrates neural networks with adaptive sliding-mode control, where a neural network dynamically learns time-varying gain matrices—eliminating requirements for system linearization, differentiability assumptions, or exact modeling. Crucially, it introduces the first end-to-end training framework that enforces physical equations as hard constraints without access to true state labels. The resulting architecture exhibits enhanced robustness against severe measurement noise, nonsmooth dynamics, and weak observability. Simulation results demonstrate rapid convergence and high-accuracy state estimation, validating its effectiveness and generalizability under complex operating conditions.

This paper addresses the robust optimal control problem for uncertain nonlinear systems. We propose a novel control architecture integrating an event-triggered mechanism (ETM), an extended state observer (ESO), and value-iteration-based adaptive dynamic programming (VI-ADP). To our knowledge, this is the first work embedding ETM directly into the ADP framework; the ESO estimates composite disturbances online, enabling disturbance compensation and closed-loop stability without requiring an exact system model. Lyapunov-based stability analysis rigorously guarantees both learning convergence and system robustness. Compared with conventional time-triggered ADP, the proposed method reduces sampling and computational load by over 60%, significantly improving computational efficiency and learning sparsity while maintaining high disturbance rejection capability and control accuracy. Numerical experiments validate the theoretical stability claims and demonstrate superior performance in terms of robustness, efficiency, and precision.

This study addresses the simulation-to-reality mismatch in robot control caused by physical disturbances despite accurate models, proposing a sampling-based disturbance observer (DOB). This method overcomes the limitation of classical DOBs that rely on explicit dynamics models by leveraging state rollout and cost query interfaces, thereby extending disturbance compensation to black-box simulators and learned world models. Furthermore, it innovatively decouples the state and cost disturbance channels for independent estimation and compensation. Experimental results demonstrate that the proposed approach effectively bridges the Sim-to-Real gap across diverse simulated and real-world robotic tasks, yielding significant improvements in control performance.

This study addresses the scalability bottleneck in certifying Lyapunov stability for large-scale neural network observers, where linear matrix inequality (LMI) constraints render semidefinite programming computationally intractable. To overcome this limitation, the authors propose a two-stage decoupled training framework that first employs point-guided pretraining to achieve high-precision local stability, followed by LMI-based fine-tuning to obtain globally rigorous stability certificates while balancing computational efficiency with safety guarantees. The work contributes theoretically guaranteed stability radii and probabilistic coverage analyses that transcend conventional scalability constraints. Compared to direct LMI approaches, the proposed method significantly accelerates training and demonstrates superior generalization capability and tracking accuracy on nonlinear control benchmarks as well as in X-29 aircraft tests.

This study addresses the challenge of composite adaptive tracking control under dynamic coupled disturbances. To this end, it proposes a predictive control framework grounded in representation learning. Specifically, the framework employs statistical methods to identify disturbance dynamics representations exhibiting contraction properties. By introducing a hard expectation-maximization algorithm augmented with a Kalman smoother, it extends conventional fixed-decay approaches into learnable predictive models and integrates Bayesian filtering for precise state estimation. Experimental evaluations on vehicles traversing slippery terrains and coupled Duffing oscillators demonstrate that the proposed method achieves accurate disturbance prediction, significantly enhancing both adaptive tracking performance and control robustness.

Traditional controllers struggle to generalize across systems with varying orders and dynamic characteristics. This work proposes a universal learning-based controller that constructs a dynamic state-space representation using a masked attention mechanism, integrating system label encoding, multi-scale temporal processing, and a mixture-of-experts architecture to enable a single neural network to uniformly control diverse linear and nonlinear systems. Notably, the approach is the first to adapt—without architectural modifications—to challenging dynamics such as unstable and non-minimum-phase systems, while supporting zero-shot generalization to unseen operating conditions. Trained on 25 system classes and 314,630 trajectories, the controller matches the performance of specialized LQI controllers and maintains robustness under previously unobserved conditions, including actuator saturation, noise, and disturbances.

This study addresses the challenge of sustained spacecraft attitude stabilization under unseen actuator faults—specifically gain errors, sign reversals, and constant biases—by introducing a novel evaluation paradigm centered on a “stability gate” metric that prioritizes long-term stability over instantaneous recovery. The proposed approach integrates recurrent neural networks for online estimation, an analytical control law, Nussbaum-gain adaptation, a disturbance observer, and multi-source sensor fusion into a structured estimation-and-control architecture. Validated on the Basilisk six-degree-of-freedom platform, the framework achieves success rates of 97.8% and 94.4% under sign-reversal and gain-fault scenarios, respectively. Notably, the inclusion of the disturbance observer elevates the success rate for constant-bias faults from 0% to 59.4%, substantially outperforming baseline methods such as end-to-end reinforcement learning and conventional PID control.

研究比较了外部、内部持续及控制生成的扰动对自适应调节的影响，发现内部持续扰动导致最大暴露和调节负担，并探讨了不同扰动源与时机如何影响模型中的暴露和控制器负担。