자율주행 배송 로봇을 위한 주행상태 적응형 다중 센서 기반 충돌 및 외부 간섭 감지

초록

Autonomous delivery robots operating in indoor and outdoor pedestrian areas must distinguish physical collisions from routine motion disturbances. Fixed-threshold inertial detectors are prone to false alarms caused by uneven surfaces, abrupt maneuvers, and wheel slip. Meanwhile, many learning-based methods rely on dedicated collision sensors or are evaluated only in constrained settings. In this paper, we present a hybrid multi-sensor collision detection framework that uses existing onboard signals. The framework's first component uses Direct LiDAR-Inertial Odometry (DLIO) estimates to evaluate command-to-motion residuals and a forward-velocity discrepancy cue against adaptive thresholds. Its second component uses a Temporal Convolutional Network (TCN) to fuse control inputs, odometry, filtered impact indicators, and a depth-based person-proximity cue. The classifier separates physical collisions from normal driving and collision-like human interference, while temporal refinement converts frame-wise predictions into stable event-level decisions. We evaluate the framework in indoor, outdoor, and mixed environments using 55 collision and 56 interference events collected across six independent sessions. The comparison with an adaptive rule-based baseline further indicates that the framework can suppress false collision decisions under the tested conditions.

키워드