This article addresses the problem of reactive collision avoidance for autonomous vehicles in environments with unknown dynamic obstacles. By developing a novel local measurement-based control barrier function (LM-CBF), a collision-free guidance module is established to make minimal modifications to the potentially unsafe commands at the kinematic level when encountering obstacles. Moreover, a sufficient condition is derived to ensure collision avoidance while considering the tracking performance of the underlying pre-equipped controllers, even in the presence of tracking errors. Compared with the prevailing schemes, the proposed method eliminates dependence on GPS, environmental models, and prior knowledge of obstacles, such as absolute position and velocity, and instead uses locally measurable distance and bearing information to react to obstacles for real-time avoidance. Furthermore, it offers superior plug-and-play functionality by neither altering the structure of the low-level tracking controller nor requiring a redesign of the kinematic guidance law. Rigorous theoretical analysis substantiates the safety properties of the proposed method. Numerical simulations and experiments on a physical robotic platform further validate its effectiveness and real-time performance in unknown environments with moving obstacles.

This paper investigates a constrained optimal control problem for safety-critical robots with parametric uncertainties. A novel adaptive safety-certified reinforcement learning (RL) algorithm is proposed, leveraging control barrier functions (CBFs) to enable safe learning of the optimal policy during the online exploration phase. Specifically, a high-order robust adaptive CBF is presented to minimally adjust RL-derived control actions by incorporating a prescribed-time adaptation law to handle the unknown system parameters. This way directly enforces forward invariance, allowing the shrunken safe set to near the standard set within a user-prescribed time. Moreover, a novel adaptive critic learning frame is presented by introducing filtered auxiliary signals that integrate both instantaneous and historical data, which relaxes the strict persistent excitation (PE) condition required in the existing RL methods to a weaker, easily verifiable finite excitation (FE) condition. Later, a prescribed-time learning rule is developed to accelerate the convergence of weights. The key advantage of the proposed way is the decoupling of safety and RL convergence, enabling each component to be managed separately, thereby offering stronger safety certifications compared to the existing RL schemes even under uncertain dynamics. The effectiveness and superiority of the proposed scheme are proven via simulations for surveillance and regulation tasks of autonomous robots.

This paper addresses an optimal elliptical enclosing problem of Unmanned Aerial Vehicles (UAVs) under prescribed constraints, whose objective is to steer UAVs to fulfill accurate target encirclement while complying with arriving time restrictions and minimum efforts. A novel learning-based approximate optimal control policy including two-stage designs is presented. At the first stage, a steady-state robust control protocol is developed to steer UAVs to precisely travel along a predefined elliptical path based on concise filtering. At the second stage, to address specified-time constraints and gain the online optimization ability, a single-critic based enhanced learning rule is explored to generate an approximate optimal regulator that stabilizes error dynamics and minimizes value functions, wherein specified-time constraints can be handled by encoding inequality conditions as skilled barrier functions, and by making full use of historical data and current information, a finite-time learning mechanism driven by weight errors rather than Bellman errors is proposed to approximate the solution of Hamilton-Jacobi-Bellman (HJB) equation with faster decaying. The distinct merit is that an improved reinforcement learning (RL) paradigm is formulated to prescribe an elliptical circumnavigation with assured time requirements and optimization behaviors, which can greatly outperform non-RL alternatives in maintaining the optimal performance index while exhibiting restriction handling ability via online learning. Lyapunov stability demonstrates that involved error variables are ultimately limited and resultant controller obeys optimality. The feasibility and values of presented algorithm are accessed by comparisons and simulations.

This paper addresses the constrained infinite-horizon optimal control problem for autonomous vehicles operating in avoidance regions. A novel online adaptive safe reinforcement learning (RL) algorithm is presented to enable real-time generation of continuous and safe motion trajectories. Specifically, the framework utilizes a safety-certified learning approach, featuring a predefined-time convergent adaptive-critic network that rapidly learns the optimal policy under mild conditions, along with a control barrier function (CBF)-based safety filter to restore original constraints through forward invariance and prevent safety violations during the online exploration phase. Rigorous theoretical analysis establishes the safety, optimality, and convergence of the RL policy. Simulations demonstrate that the proposed scheme effectively generates safe, near-optimal trajectories for autonomous navigation tasks, with comparative evaluations highlighting its superiority in optimizing long-term performance over the prevailing motion planners with obstacle avoidance, while maintaining competitive execution time.

The target enclosing control problem for autonomous vehicles with uncertainties necessitates simultaneous consideration of control optimality, robustness, and safety-guided performance constraints. This paper presents a performance-prescribed optimal control algorithm using control barrier function (CBF)-based reinforcement learning (RL) to address the above problem, which contains two key contributions. First, a special CBF-based argument term is developed and embedded into the reward function to characterize environmental feedback regarding the risk of violating constraints, which enables the controller to confine enclosing errors within declared boundaries with minimal intervention. Second, a critic-only neural network is utilized to synthesize the optimal control policy, where a novel fixed-time updating law is presented to accelerate the weight convergence to ideal values within a fixed settling time, thereby enhancing the online learning ability and further improving control performance. Theoretical outcomes related to learning convergence, safety, stability, and robustness are rigorously verified. Simulations reveal that the proposed strategy outperforms the previously designed enclosing controllers based on the non-RL and RL ways in terms of complying with prescribed safety constraints and optimizing long-term performance.

This article studies the pursuit–evasion game involving nonholonomic vehicles constrained by input saturation, aiming for the pursuer to intercept an evasive opponent. Unlike the previous game research neglecting the practical kinematic constraints, a coupled nonlinear system is formulated to elucidate the interaction dynamics between the players. After that, the optimal control strategies are derived by solving the Hamilton–Jacobi–Isaacs (HJI) equation linked to a special nonquadratic cost function. The Nash equilibrium analysis and finite-time capturability are conducted. To learn the optimal pursuit–evasion strategy pair, a fixed-time convergent reinforcement learning (RL) algorithm is proposed, which leverages a novel residual design to facilitate weight updates by collecting and evaluating current and historical data based on information quality. Compared with the existing RL methods that suffer from sluggish convergence due to an asymptotic learning rule and the stringent persistent excitation (PE) condition, the proposed RL relaxes the PE to an easily achievable and online verifiable finite excitation (FE) condition, allowing rapid weight convergence within a fixed period. Simulations and comparisons validate the effectiveness and superiority of the proposed method, showing a 61% reduction in convergence time in contrast to the prevailing RL schemes.

This article investigates a finite-time composite learning-based elliptical enclosing control for nonholonomic robots under a global positioning system (GPS)-denied environment. At the kinematic level, following a prediction and innovation architecture, a novel bearing measurement-based relative position observer formulated in a local coordinate is proposed to assure an exponential decaying of estimation errors without the aid of GPS. Utilizing the observation outcomes, an elliptical guidance law with a time-varying enclosing radius and nonorthogonal tangential and axial vectors is established to yield the reference velocity and angular rate to be tracked. At the kinetic level, by constructing filtering operations and auxiliary variables to extract weight errors, a special finite-time composite neural learning driven by weight and tracking errors is devised to reinforce parameter convergences, then an anti-disturbance kinetic control rule is designed to achieve online precise disturbance compensation and finite-time error convergence. The distinct merit is that an elliptical surrounding concerning an unknown target can be fulfilled with the finite-time neural learning capability while eliminating the deployment of GPS, which is nontrivial and challenging than reported circumnavigation alternatives either relying on the accessibility of GPS or neglecting kinetic uncertainties. An input-to-state stable criterion is applied to demonstrate the boundedness of a closed-loop system. Simulations are provided to confirm the utility of the considered strategy.