Conference Agenda

As conflicts between wildlife and humans have become an increasing problem, mobile robots that deter wildlife by emitting human voices or predator sounds while moving around are expected to provide deterrent effects over wide areas while suppressing habituation. This study aims to develop an autonomous mobile platform capable of continuous operation in GNSS-denied environments such as forest work roads.

We developed a robot localization system based on loosely coupled 3D LiDAR and IMU, incorporating point cloud distortion correction and a recovery algorithm for localization failure. A navigation system was also implemented to follow predefined routes using pure-pursuit control, detect obstacles and stop safely, and autonomously dock with a charging station. Driving experiments were conducted in a tree-dense environment with uneven terrain on a university campus, and the root mean square error (RMSE) between the planned route and the actual trajectory was evaluated.

Twelve driving trials were performed, and the mean RMSE was 3.6 cm. Autonomous docking was successful in all trials, and the robot was able to safely stop when a human blocked the route.

These results indicate that the developed platform can achieve stable, continuous autonomous operation in forest-like environments and is applicable to wildlife damage mitigation.

This study proposes a vision-based navigation pipeline for a rice weeding robot that predicts steering angles from crop structure while preserving robustness in dynamic paddy fields. We employ YOLOv8seg to perform root-level instance segmentation of seedlings and fuse masks across frames to derive a stable row-direction prior that directly drives steering angle prediction and lateral guidance. To address visual localization failures caused by foliage motion, water-surface specularity, and repetitive texture, we tightly couple semantic masking with ORB‑SLAM3 and replace fragile keypoint detectors with mask-driven three-dimensional root anchors obtained by neighborhood depth aggregation and temporal smoothing. Preliminary verification indicates improved tracking continuity and accurate row‑heading estimation under variable wind and illumination. In the future, we will complete ablation studies comparing semantic masking, single versus multi-frame fusion, and monocular with learned depth priors versus stereo or RGB-D sensing, and we will conduct closed-loop evaluations using LQR (Linear Quadratic Regulator), Stanley, and PID (Proportional Integral Derivative) controllers. The results aim to demonstrate that object level agronomic semantics combined with geometric priors enable reliable and interpretable steering for crop-scale field robotics and provide a reproducible pipeline for structure‑aware navigation.

As labor shortages increasingly limit horticultural greenhouse harvesting, autonomous robots are emerging as a solution – yet enabling real-time multi-camera perception on edge devices demands robust, resource-efficient processing. Specifically, executing concurrent perception tasks – such as object detection, depth estimation, and localization – on battery-powered platforms poses significant challenges in balancing computational load against power consumption. To address these constraints, we deployed a multi-camera perception pipeline for a quadruped Korean melon harvesting robot built on the Ghost Robotics Vision 60 platform, equipped with a ZED2i stereo camera and two ArduCam modules via an NVIDIA Jetson AGX Orin. We systematically evaluated Deep Learning Accelerator (DLA) utilization against GPU processing in power consumption, GPU workload, and latency. Our results demonstrate that DLA offloading substantially improves power efficiency, reducing consumption by up to 82.7% for YOLO11x and lowering GPU utilization by 74.4 percentage points (from 77.5% to 3.1%). This releases substantial GPU capacity for other onboard tasks such as SLAM, motion planning, and visual servoing. Additionally, INT8 quantization for DLA inference improved processing speeds by up to 33.7%, meeting real-time requirements. These findings provide empirical evidence that hardware-level optimizations offer a practical pathway toward deploying robust perception systems in real-world autonomous agricultural applications.

Autonomous agricultural robots commonly rely on single-antenna GNSS receivers due to cost constraints. However, such systems suffer from severe heading instability at low speeds and pose discontinuities during GNSS signal interruptions and subsequent reacquisition. This paper proposes a lightweight localization and heading stabilization framework tailored for low-cost platforms. The method combines kinematic prediction with event-driven GNSS updates, velocity-scheduled heading stabilization, and a recovery smoothing strategy. A low-cost IMU is incorporated only for short-term angular velocity propagation to maintain motion continuity during GNSS degradation. Unlike tightly coupled GNSS/INS systems, the proposed approach avoids complex heading stabilization and bias estimation, making it suitable for real-world agricultural deployment. The effectiveness of the method is validated in a closed-loop navigation context. Simulation experiments demonstrate reduced heading variance at low speeds and smoother trajectory recovery during GNSS outages.

This study investigates the effects of grouser shape and contact angle on lateral slip reduction in a slope-adaptive crawler mower equipped with a center-of-gravity shifting mechanism. Previous experiments confirmed that this mechanism reduces slip; however, the contribution of grouser cross-sectional shape remains unclear. Therefore, this study focuses on optimizing grouser geometry and contact conditions. Laboratory experiments using several grouser shapes were conducted on two soil types under different loading conditions, and dynamic friction coefficients were calculated. In addition, discrete element method (DEM) simulations were performed to evaluate both dynamic and maximum static friction coefficients under contact angles from 0° to 70°. Experimental results showed no significant differences among grouser shapes on hard soil, whereas on softer soil, inclined angular grousers exhibited relatively higher friction. Simulation results indicated that arch-shaped grousers achieved the highest friction. Furthermore, increasing the contact angle enhanced friction due to increased uphill force components and soil accumulation. These findings suggest that optimizing grouser shape and controlling contact angle can significantly improve slope stability.

Accurate estimation of fruit pose in three-dimensional space is a critical prerequi-site for autonomous fruit grasping. This study proposes a rotation pose estima-tion method for Ichida persimmons based on RGB and depth images. In this study, the rotation of a persimmon is defined as a direction vector from the geo-metric center of the fruit to the base of the calyx. To obtain large-scale annotated training data, a synthetic data generation approach based on persimmon point clouds is developed, which enables rapid rendering of images and corresponding labels under diverse poses. The pose estimation model is constructed and trained using ResNet34 as the backbone. In the experiments, synthetic data generated from 100 persimmons were used for training, while a real-image dataset consist-ing of 227 images from 20 persimmons was employed for testing. The results show that the model achieves a mean angular error of 4.89°, a root mean squared error of 5.52°, and a standard deviation of 2.55° on the test set, with 97.80% of samples exhibiting errors below 10°. Although a noticeable sim-to-real gap re-mains, this study demonstrates the feasibility of synthetic data–driven fruit pose estimation and provides a foundation for future research integrating more ad-vanced simulation and real-world applications.

The performance of current robots in unstructured greenhouse environments still remains limited due to large variability in lighting conditions, irregular plant shapes, and severe occlusions, which result in missed detections of plant parts, false-positive detections, and prolonged observation times to obtain a 3D plant reconstruction for downstream tasks such as harvesting or de-leafing. Active-vision methods mitigate these challenges by selecting the next-best view (NBV) to enhance the reconstruction. However, current active vision methods still fall short in efficiency.

In this work, we enhance NBV planning through two complementary methods. First, we extend NBV planning with a main-stem-guided sampling strategy, which samples candidate viewpoints around the main stem and the next expected node. Second, we introduce a multi-view filtering method that reduces false-positive node reconstructions by verifying consistency across viewpoints, while accounting for occlusions. The proposed methods were implemented and evaluated on a robotic arm equipped with an RGB-D camera. Experimental results for 3D node reconstruction in a tomato greenhouse demonstrate a 71% reduction in false positives and a 9% reduction in the number of required viewpoints, compared to baselines without the proposed enhancements. These improvements enhance the feasibility of robotic harvesting and de-leafing in commercial greenhouse environments.