Obstacle Modeling and State Estimation for Mobile Robots Toward Real-World Deployment of Non-Reciprocal ORCA

Abstract

Non-reciprocal ORCA is an efficient velocity-space method for local collision avoidance, but its practical deployment usually assumes that obstacle position, size, velocity, and motion category are already known. In real robotic systems, however, such information is rarely available directly from raw sensor observations. To address this issue, this paper proposes an obstacle modeling and state estimation method for the real-world deployment of non-reciprocal ORCA. Based on 2-D LiDAR point clouds, circular obstacle primitives are first generated by direct point-wise mapping and then sparsified by removing redundant circles under a coverage constraint. Visual semantic priors and spatial connectivity are further combined to distinguish static and dynamic obstacles, while cluster-level Kalman filtering is adopted to estimate the motion states of dynamic obstacle clusters. The resulting obstacle position, radius, and velocity are finally provided to non-reciprocal ORCA as unified inputs. Simulation and real-world experiments demonstrate that the proposed method improves obstacle representation quality, trajectory smoothness, and avoidance efficiency in both static and dynamic scenarios.