Abstract: To address incomplete environmental perception, excessive trajectory deviation, and insufficient post-obstacle-avoidance recovery of agricultural robots operating in narrow greenhouse corridors, this study aimed to develop a local obstacle avoidance method integrating multi-sensor perception with an improved timed elastic band (TEB) planner. The method was intended to improve low-lying obstacle perception, centerline keeping, boundary safety, and pose recovery after obstacle avoidance.A two-dimensional light detection and ranging (2D LiDAR) sensor and a red-green-blue-depth (RGB-D) camera were fused to construct corridor boundaries, the center cruise line, and local obstacle representations. The LiDAR provided stable planar boundary information, whereas the RGB-D camera supplemented short-range and low-lying obstacle information. Multi-source point clouds were transformed into a unified robot coordinate system and projected onto the local plane. A grid-based occupancy representation was adopted to reduce the influence of point-cloud density differences, and a local risk model was established to characterize obstacle intrusion into the effective passage region. Greenhouse corridor geometric guidance constraints and recoverability constraints were then incorporated into the conventional TEB framework. The former penalized excessive centerline deviation and insufficient boundary clearance, whereas the latter penalized terminal position and heading errors. A cruise-obstacle avoidance-recovery/alarm state-switching strategy was also designed.The proposed method was evaluated in four two-dimensional simulation scenarios, including single-obstacle avoidance, a 5 m constrained corridor, an asymmetric narrow passage, and continuous staggered obstacles, and was further verified through real-vehicle experiments in a greenhouse. The dynamic window approach (DWA), artificial potential field (APF), conventional TEB, and improved TEB were compared using success rate, average planning time, minimum obstacle clearance, minimum wall clearance, maximum centerline offset, and terminal heading error. In the single-obstacle scenario, the improved TEB achieved a 100% success rate, a maximum centerline offset of 0.7537 m, and a terminal heading error of 0.0017 rad. In the constrained corridor scenario, it achieved a 97% success rate and a terminal heading error of 0.0157 rad, compared with 94% and 0.2361 rad for the conventional TEB. In the asymmetric narrow-passage scenario, the improved TEB achieved a 99% success rate, a minimum wall clearance of 1.7113 m, a maximum centerline offset of 0.2994 m, and a terminal heading error of 0.0008 rad. In the continuous staggered-obstacle scenario, it achieved a 96% success rate, reduced the maximum centerline offset from 1.2795 to 0.6175 m, increased the minimum wall clearance from 0.5095 to 1.2750 m, and reduced the terminal heading error from 0.6244 to 0.0015 rad relative to the conventional TEB. Ablation experiments showed that the geometry-guidance-only variant reduced the maximum centerline offset from 1.2795 to 0.6042 m and increased the minimum wall clearance from 0.5126 to 1.2944 m, whereas the recoverability-only variant reduced the terminal heading error from 0.6314 to 0.0007 rad. With both constraints enabled, the complete improved TEB achieved a maximum centerline offset of 0.5997 m, a minimum wall clearance of 1.3052 m, and a terminal heading error of 0.0012 rad. In real greenhouse experiments, the improved TEB achieved a 93.3% success rate. Compared with the conventional TEB, it reduced the maximum centerline offset from 0.768 to 0.349 m, increased the minimum wall clearance from 0.728 to 0.905 m, and reduced the terminal heading error from 0.0416 to 0.0013 rad. The method also completed obstacle avoidance under regular, low-lying, and moving obstacle conditions and switched to safe stopping when no feasible passage satisfied the safety constraints.The proposed method provides an effective local obstacle avoidance solution for agricultural robots in narrow greenhouse corridors. It improves the complementary use of heterogeneous sensor information, strengthens centerline maintenance and boundary safety, and enhances post-obstacle-avoidance pose recovery while maintaining a high obstacle-avoidance success rate. The method is suitable for continuous autonomous operation in constrained greenhouse environments and provides a technical reference for local environmental perception and path planning.