Aug 2026· IEEE Transactions on Neural Networks and Learning Systems· Vol PP· 0 citations
Medicine
TL;DR
The Safe-Koopman Framework is introduced, an operator-theoretic motion planning method that generalizes obstacle-free demonstrations to planar planning tasks with obstacles and avoids collisions observed in the unconstrained Koopman baseline.
Abstract
Robot learning from demonstration (LfD) in obstacle-laden environments presents a fundamental conflict between mimicking expert dynamics and maintaining safety. This article introduces the Safe-Koopman Framework, an operator-theoretic motion planning method that generalizes obstacle-free demonstrations to planar planning tasks with obstacles. Leveraging the Koopman operator with random Fourier features (RFFs), the method lifts nonlinear robot motion into a linear feature space to capture the global kinematic topology of demonstrations. To address safety generalization, two components are proposed. First, a dynamical system modulation mechanism smoothly blends the nominal imitation flow with a local safety-critical flow, reducing discontinuities associated with hard switching. Second, a radial loss is formulated as a data-driven relaxation of control barrier function (CBF) boundary conditions, penalizing inward velocity components near obstacle boundaries. Under the zero-loss idealization and stated regularity assumptions, the radial condition supports forward invariance of the safe set; in implementation, the method empirically maintains positive clearance. The nominal spectral stability of the learned Koopman operator is further analyzed, along with the effect of local nonlinear modulation on this conclusion. Experiments on the LASA handwriting trajectory benchmark show that the proposed method avoids collisions observed in the unconstrained Koopman baseline and reduces jerk compared with artificial potential field (APF)-based obstacle avoidance while maintaining competitive kinematic fidelity.
A task-space receding-horizon controller that uses a short contact-consistent rollout to generate a terminal kinematic reference satisfying internal non-penetration constraints, then computes only the first input of a smooth minimum-acceleration transition toward that reference.
Collaborative robots (cobots) operating in shared human-robot workspaces require reactive, computationally lightweight obstacle avoidance to guarantee both safety and task continuity. This paper presents a simulation-based study of obstacle avoidance for the 6-DOF UR3 cobot (Universal Robots) using Modulation Dynamical System approach (MDS). A nominal autonomous dynamical systems (DS) with a stable target attractor is first designed to generate goal-directed endeffector motion. Reactive collision avoidance is then achieved by locally modulating the nominal velocity field through a modulation matrix constructed from each obstacle's geometric representation, deflecting the flow around obstacles while preserving asymptotic convergence to the attractor and avoiding the need for global path re-planning. The framework is evaluated in simulation across scenarios containing single and multiple static obstacles, where the cobot consistently reaches the target along collision-free trajectories. Limitations of the approach are analyzed and discussed, motivating directions for future work.
Le Nam Chau, Pham Vo Khai Anh, Nguyen Thi Phuong Mai et al.· 2026 11th International Conf...· 0 citations
Addressing the challenges of obstacle avoidance for autonomous vehicles in complex dynamic environments, traditional Artificial Potential Field (APF) methods often suffer from delayed responses to dynamic obstacles and generate paths that violate vehicle kinematic constraints. To overcome these limitations, this paper proposes an improved path planning algorithm, the Relative Velocity Potential Field (RVPF), which fuses relative velocity information with kinematic constraints. First, a relative velocity sensitivity factor is introduced to construct a dynamic potential field. By dynamically reshaping the repulsive field distribution based on the relative velocity vector, this approach endows the algorithm with a predictive capability regarding collision risks, facilitating a transition from passive reaction to active defense. Second, a vehicle kinematic model is established incorporating Ackermann steering geometry. A virtual tangential force strategy is employed to map the resultant potential forces into control variables that adhere to wheelbase and steering angle limits, thereby ensuring the generation of smooth and feasible trajectories. Simulation results demonstrate the superior performance of the proposed algorithm in scenarios involving high-speed oncoming traffic, overtaking, and lateral crossing. Notably, in the lateral crossing scenario—where traditional APF failed due to collisions—the RVPF algorithm achieved collision-free passage by actively decelerating and yielding, increasing the minimum safety distance to 5.02 m. These results confirm that the proposed algorithm significantly enhances the safety and stability of autonomous vehicles across diverse traffic situations.
The experimental results demonstrate that the proposed Field-guided RRT achieves a better balance among path efficiency, planning time, obstacle-clearance maintenance, and trajectory execution capability compared with conventional RRT-based methods.
Chen-Yueh Wen, Siyuan Zhang, Maksim A. Grigorev et al.· Machines· 0 citations
Local navigation in complex static obstacle environments requires an unmanned aerial vehicle to reach a target while avoiding obstacles under limited local perception. U-shaped non-convex obstacles acting as local traps further increase the difficulty of detour selection and continuous-control execution. A key challenge is to jointly capture forward scene structure, surrounding geometric constraints, and execution-level control requirements within a unified decision-making pipeline. To address this challenge, we present PC-MM-SAC, a task-oriented multimodal reinforcement-learning framework with a lightweight physics-constrained action-execution layer for local navigation in complex static obstacle environments. The method learns from structured observations built from depth maps, LiDAR measurements, and task-related state variables. The layer applies bounded action mapping and yaw-rate variation limiting to improve the smoothness and continuity of continuous-control commands without introducing an additional online optimization-based controller. We further employ an event-conditioned modality value-sensitivity analysis during evaluation to characterize how the critic’s local sensitivity to different information sources varies across representative navigation phases. In AirSim experiments, PC-MM-SAC achieves a success rate of 0.88 and an episode reward of 772, attaining the highest success rate among the evaluated methods under the current task setting. Ablation and behavioral analyses suggest that multimodal observations are the primary contributor to the observed performance improvement, while the physics-constrained action-execution layer is associated with smoother trajectory execution and reduced local oscillation.
Yawei Tian, Yu Chen, Zhongliang Deng et al.· Scientific Reports· 0 citations
This follow-up work tests the feasibility of the neuro-inspired self-supervised learning framework for trajectory planning that leverages forward and inverse models as the internal supervisory mechanism in an environment that contains an obstacle, and demonstrates the tendency of the planner to exploit the learning signal provided by the forward and inverse models.
Miroslav Krupa, Miroslav Cibula, Kristína Malinovská· 0 citations