As AI systems increasingly assist humans in physical tasks, ensuring safety becomes paramount – physical actions carry immediate and irreversible consequences that digital errors do not. We introduce the Vision-Language Embodied Safety Agent (VLESA), a framework that monitors human activities from egocentric video and triggers real-time safety interventions when dangerous actions are predicted. VLESA addresses intent-dependent safety where identical actions can be safe or dangerous depending on context. A dataset pairing egocentric frames with goal-conditioned safety annotations is introduced, enabling a goal-conditioned safety Q-filter trained via GRPO that evaluates actions with respect to inferred intent without retraining. On top of that, an intent-action prediction agent is proposed to jointly infer goals and predict future actions from video. On the ASIMOV-2.0 benchmark, VLESA achieves higher intervention accuracy at the exact ground-truth frame compared to baselines, while the GRPO-trained Q-filter improves action safety by over 41 percentage points through goal-conditioned constrained decoding. Code is available at https://github.com/HanjiangHu/VLESA.

Reactive task-space planners such as Bug2 operate with fixed Cartesian step sizes and are unaware of the manipulator's joint-angle limits. When the Jacobian is poorly conditioned, even small Cartesian steps can demand joint changes that exceed admissible bounds; clipping the joints to their limits causes tracking drift and can prevent goal reaching entirely. We address this by computing, at each planning step, the largest Cartesian hyperrectangle that is certifiably reachable under joint displacement bounds. Using a second-order polynomial approximation of the inverse kinematics and the S-procedure, we formulate a small semidefinite program whose solution yields the certified half-width λ^⋆. An equivalent bisection procedure exploiting the quadratic structure solves the certification in sub-millisecond time. Integrating this certificate with Bug2 yields a planner whose step size adapts to local kinematic conditioning. In a statistical evaluation over 94 adversarial scenarios spanning six joint-limit settings, the SOS-verified planner achieves zero joint-limit violations with a 100% goal-reaching rate, whereas a standard Bug2 planner violates joint limits in 6–11% of steps and fails to reach the goal in up to 18% of scenarios.

In this paper, we develop a novel solution to the output-feedback adaptive optimal control problem of general nonlinear nonaffine systems based on reinforcement learning (RL). This bridges a gap between RL and nonlinear output-feedback control theory by designing high-fidelity data-driven controllers to achieve disturbance rejection in an optimal sense. Based on the condition of uniform observability, the state is reconstructed by the retrospective input and output information, which can be seen as equivalent to a deadbeat observer. Both policy and value iteration algorithms are proposed to learn the optimal output-feedback control policy and value function. The convergence of the proposed algorithms and the practical stability of the closed-loop system with learned control policy are rigorously ensured even when the optimal value function is not positive definite. (c) 2026 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

The co-design paradigm claims substantial advantages to hardware and control system design by addressing multidisciplinary challenges within a unified framework. Established co-design frameworks for robot manipulators have predominantly focused on two components: motor/arm design and trajectory optimization, which inadequately address real-world disturbances and model uncertainties and thus render suboptimal design and closed-loop system performance. This paper proposes a closed-loop co-design (CLCD) framework to jointly determine motors, motions, and a feedback controller, where the controller comprises a reinforcement learning (RL)-based compensator and a classic proportional-derivative controller for trajectory tracking. Simulation is performed to validate (1) the effectiveness of the proposed CLCD framework to attenuate the sim-2-real gap, (2) the viability of incorporating an RL-based controller into the CLCD for flexible and efficient synthesis of control policy, and (3) the scalability of the CLCD by applying it to perform co-design for 12 and 120 tasks.

Reinforcement learning (RL), as an important branch of machine learning, has recently achieved extensive attention and success in many applications. Its main idea is to enable agents to continuously learn to make optimal decisions by trying to maximize a reward function for their actions and interactions with the environment. However, making high-quality decisions in complex and uncertain real-world scenarios is a challenging task. The interference and attacks in such scenarios tend to destroy the existing strategies. Maintaining RL's optimal performance in various cases and adapting to changing environments remains an important challenge. This article presents a comprehensive review of recent advancements in robust reinforcement learning (RRL), and analyzes them from the perspectives of challenges, methodologies, and applications. It systematically evaluates current progress in RRL and summarizes the commonly used benchmark platforms. Finally, several open challenges are discussed to stimulate further research and guide future developments in this area.

Electrified high-speed train transportation systems utilize inverter-driven traction motors and a variable-voltage variable-frequency (VVVF) control scheme to achieve high efficiency across a wide range of speeds. However, due to the varying operating conditions, it becomes challenging to monitor the health of traction motors, especially when the train voltage and speed are unknown or not accurately measured. In this paper, we address the problem of extracting current signatures from train traction motors for fault detection without knowing the instant voltage amplitude and the instant motor speed. We propose a robust algorithm that combines minimum-variance spectral analysis to mitigate the effects of varying voltage and a speed compensation technique to tackle the varying-speed issue for extracting fault signatures in the stator current. Experimental results on different current data, including faulty laboratory motor current, onsite measurements from healthy train motors, as well as synthesized faulty train motor current, show that our method can accurately estimate the motor speed and achieve a robust spectrum under various operating conditions for fault signature extraction.

This paper studies system identification for nonlinear state-space models, a problem that arises across many fields yet remains challenging in practice. Focusing on maximum likelihood estimation, we employ Bayesian optimization (BayesOpt) to address this problem by leveraging its derivative-free global search capability enabled by surrogate modeling of the likelihood function. Despite these advantages, standard BayesOpt often suffers from slow convergence, high computational cost, and practical difficulty in attaining global optima under limited computational budgets, especially for high-dimensional nonlinear models with many unknown parameters. To overcome these limitations, we propose an accelerated BayesOpt framework that integrates BayesOpt with the Nelder–Mead method. Heuristics-based, the Nelder–Mead method provides fast local search, thereby assisting BayesOpt when the surrogate model lacks fidelity or when over-exploration occurs in broad parameter spaces. The proposed framework incorporates a principled strategy to coordinate the two methods, effectively combining their complementary strengths. The resulting hybrid approach significantly improves both convergence speed and computational efficiency while maintaining strong global search performance. In addition, we leverage an implicit particle filtering method to enable accurate and efficient likelihood evaluation. We validate the proposed framework on the identification of the BattX model for lithium-ion batteries, which features ten state dimensions, 18 unknown parameters, and strong nonlinearity. Both simulation and experimental results demonstrate the effectiveness of the proposed approach as well as its advantages over alternative methods.

Hyper-redundant manipulators (HRMs) are used in noninvasive on-site maintenance tasks of aero-engines owing to their excellent dexterity and adaptability. However, it is challenging for them to achieve the obstacle avoidance planning in confined environment compared to conventional manipulators. This work presents a 3-D-dynamic safety envelope (3DSE)-based collision avoidance strategy to tackle this challenge. First, unlike conventional obstacle expansion-based methods that are computationally expensive and conservative, 3DSE is proposed and applied to establish collision constraints. Then, the motion in confined spaces is divided into trajectory tracking and blade detection processes. To address the issue of HRM’s inability to fully track the trajectory, the continuous motion trajectory is discretized, and 3DSE is employed to update the trajectory nodes. The blade detection process when the base is fixed is modeled as an inverse kinematics (IK) problem of HRM, and a heuristic serpentine scanning method based on 3DSE and regional adaptation is proposed. An 18 degrees-of-freedom HRM prototype is built to experimentally validate the mechanical design and the performance of the proposed method. The results show that it well completes the traversing and fault detection of multistage blades in confined space.

This paper addresses real-time motion planning for servo systems subject to nonlinear, state-dependent actuator constraints. A reference reshaping method is proposed that combines analytical optimality with high computational efficiency. Using Karush-Kuhn-Tucker conditions, the problem structure is explicitly characterized, and it is shown that the optimal solution lies within a finite set of candidate points. The complete solution set is constructed via closed-form expressions and a small-scale eigenvalue problem, yielding a deterministic algorithm that recovers the exact optimal solution without iterative optimization or root-finding. To address kinematic mismatch induced by aggressive commands, a real-time trajectory compensator is introduced to correct accumulated position error while preserving feasibility. Simulation results demonstrate significant computational speed improvements over existing methods, enabling real-time implementation at frequencies approaching 100 kHz.

Large Language Models (LLMs) have demonstrated remarkable capabilities in knowledge acquisition, reasoning, and tool use, making them promising candidates for autonomous agent applications. However, training LLM agents for complex multi-turn task planning faces significant challenges, including sparse episode-wise rewards, credit assignment across long horizons, and the computational overhead of reinforcement learning in multi-turn interaction settings. To this end, this letter introduces a novel approach that transforms multi-turn task planning into single-turn task reasoning problems, enabling efficient policy optimization through Group Relative Policy Optimization (GRPO) with dense and verifiable reward from expert trajectories. Our theoretical analysis shows that GRPO improvement on single-turn task reasoning results in a lower bound of the multi-turn success probability under the minimal turns, as well as the generalization to subtasks with shorter horizons. Experimental evaluation on the complex task planning benchmark demonstrates that our 1.5B parameter model trained with single-turn GRPO achieves superior performance compared to larger baseline models up to 14B parameters, with success rates of 70% for long-horizon planning tasks.

This article investigates real-time fast charging for lithium-ion batteries at high C-rates. First, an electrochemical-thermal-inspired battery model is presented to capture key dynamics accurately. The model is validated through hardware experiments, showing small modeling errors across a wide range of charging (up to 4C) and discharging (up to 14.5C) currents. A fast-charging framework is then proposed, consisting of two components: an offline trajectory optimization and an online current reshaping algorithm. First, the offline component solves a time-optimal charging trajectory optimization problem once at the start of the charging process. This generates an optimal reference trajectory for battery states and controls, which is then used by the online component. Second, the online component continually reshapes the reference charging current in real time. Operating at a higher frequency, it adjusts the current based on the present state and the reference current for the next time instance, ensuring compliance with charging constraints. Numerical experiments confirm the effectiveness and computational efficiency of the proposed framework.

This paper proposes a systematic procedure to enhance the torque capability of sensorless induction machine control at/near zero frequency/speed under high-frequency signal injection (HFSI). The key idea is to select the tilted injection angle by two criteria: (1) maintaining the polarity of the sensitivity of an error signal w.r.t. the rotor flux angle estimation error to ensure closed-loop stability, and (2) maximizing the sensitivity to ensure robustness against uncertainties. Three look-up tables of injection angles and error scalings are constructed regarding operating points and used for real-time HFSI implementation. We further integrate the proposed HFSI method with an adaptive flux observer to enable a stable sensorless torque control at full-speed range. Experimental results validate the torque capability and demonstrate smooth operation over the full speed range of the proposed unified observer.

This paper introduces a hierarchical motion planning strategy for autonomous tractor-trailer systems, designed for efficient long-horizon, collision-free maneuvering in complex environments. By combining high-level reference line graph search with low-level primal-dual reinforcement learning (RL)-based trajectory optimization, our approach addresses the computational challenges inherent to the motion planning of tractor-trailer dynamics. The high-level graph search decides waypoints guided by Reeds-Shepp cost, and the low-level RL connects the waypoints with dynamically feasible and collision-free trajectories. To enhance safety and accuracy, we incorporate reachability constraints and batch trajectory sampling in the RL algorithm design. Empirical results show that our method significantly reduces computation time, outperforming traditional state-lattice-based planning approaches and enabling real-time applicability.

Large-scale battery energy storage systems (BESSs) are accelerating the clean energy transition and enhancing energy resilience. They require optimal power management (OPM) to minimize power losses under various operational constraints by optimally allocating power among constituent units. Conventional formulations involve large-scale nonlinear nonconvex constrained optimization, which is computationally intractable and suffers performance degradation under uncertain power demands. In a departure, this article inaugurates a new framework to address these bottlenecks-we introduce power-sharing ratios (PSRs), which specify each cell's power quota from the output power demand, and propose to infer their values using the BESS's optimal future behavior as the evidence. To furnish the framework, we present an OPM formulation based on nonlinear model predictive control (NMPC), along with a parameterized representation of PSR; then, we translate this formulation into a Bayesian inference problem to estimate PSR-related parameters; finally, we leverage the ensemble Kalman inversion technique to achieve parameter estimation. The proposed framework combines PSR, parameterization, and Bayesian parameter estimation to scale down the OPM problem while solving it with high computational efficiency. The solution is flexible for implementation under uncertain power demands. We validate the effectiveness of the proposed framework through extensive simulations and experiments on a 20-cell BESS prototype.

This paper proposes a novel unified observer that combines a modified low-frequency signal injection (LFSI) estimator and an adaptive full-order (AFO) flux observer for smooth sensorless control of nonsalient induction machines at full speed range. The unified observer adopts the same architecture as the AFO, i.e., comprising a rotor speed estimator and a flux observer. The speed estimator dynamics are enforced by an error signal from either the modified LFSI or the AFO, based on a switching signal indicating the system’s observability at an operation point. In order to achieve a smooth transition during the switch, the LFSI channel is modified to enable the switching on the speed estimate derivative rather than the conventional approach of switching on the speed signal. A thorough theoretical analysis is provided. Simulations and experiments validate that the proposed unified observer results in smooth operation over the full speed range.

This paper proposes a novel physics-informed ma-chine learning framework for motion planning and control of autonomous vehicles. By integrating longitudinal and lat-eral control, a nonlinear control problem is formulated using Model Predictive Control (MPC). To address computational challenges, a self-supervised framework, Recurrent Predictive Control (RPC), is introduced, leveraging differentiable neural networks and recurrent neural networks to train a neural network controller. Additionally, a heuristic feedback control layer is designed to reduce steady-state errors in the closed-loop tracking. Through numerical simulations and co-simulations using Simulink and CarSim, five neural network controllers are compared with an MPC controller in a lane-changing sce-nario. The proposed RPC framework improves computational efficiency by 95% compared to MPC, enhances generalization performance compared to Approximate MPC, and reduces performance loss by 17% compared to Differentiable Predictive Control. The heuristic feedback control layer further reduces steady-state errors and improves convergence speed during training.

Electric vertical take-off and landing (eVTOL) aircraft have emerged as a promising solution to transform urban transportation. They present a few technical challenges for battery management, a prominent one of which is the prediction of the power capability of their lithiumion battery systems. The challenge originates from the high C-rate discharging conditions required during eVTOL flights as well as the complexity of lithium-ion batteries’ electrothermal dynamics. This paper, for the first time, formulates a power limit prediction problem for eVTOL which explicitly considers long prediction horizons and the possible occurrence of emergency landings. We then harness machine learning to solve this problem in two intertwined ways. First, we adopt a dynamic model that integrates physics with machine learning to predict a lithium-ion battery’s voltage and temperature behaviors with high accuracy. Second, while performing search for the maximum power, we leverage machine learning to predict the remaining discharge time and use the prediction to accelerate the search with fast computation. Our validation results show the effectiveness of the proposed study for eVTOL operations.

This paper presents a framework for the planning and control of wheeled mobile manipulators, integrating a novel reduced-order dynamics model that can be used during motion planning to enhance safety and prevent tip-overs. Leveraging centroidal momentum dynamics, the model captures key forces at the center of mass, enabling efficient, dynamically feasible trajectory generation that considers both manipulator motion and ground interaction forces. Unlike traditional methods that rely on conservative planning or separate reactive tip-over prevention mechanisms, our approach incorporates stability considerations directly into the motion planning phase. By embedding the zero moment point criterion within the model, our framework ensures tip-over prevention even during fast and payload-intensive tasks. Simulations demonstrate the effectiveness of this approach, achieving stable and efficient task execution across various scenarios.

In this paper we address the simultaneous collision detection and force estimation problem for quadrupedal locomotion using joint encoder information and the robot dynamics only. We design an interacting multiple-model Kalman filter (IMM-KF) that estimates the external force exerted on the robot and multiple possible contact modes. The method is invariant to any gait pattern design. Our approach leverages pseudo-measurement information of the external forces based on the robot dynamics and encoder information. Based on the estimated contact mode and external force, we design a reflex motion and an admittance controller for the swing leg to avoid collisions by adjusting the leg's reference motion. Additionally, we implement a force-adaptive model predictive controller to enhance balancing. Simulation ablatation studies and experiments show the efficacy of the approach.