This work focuses on grasp planning using a humanoid robotic hand. It explores reinforcement learning, imitation learning, and learning from demonstration to achieve dexterous and human-like manipulation. Experiments are conducted in both IsaacLab simulations and real-world environments. The goal is to enable robust grasping and manipulation in unstructured environments, such as bin picking of non-standardized objects.
The underwater domain presents challenges due to the lack of large-scale datasets and the high costs associated with field experiments. To address this, I have developed a ROS Gazebo-based underwater simulator and proposed several techniques for sim-to-field domain transfer including sim-real style transfer, noise randomization, and 3D medium-based adaptation. As a result, a deep learning model, trained only in simulation, was deployed on an autonomous underwater vehicle, successfully detecting and tracking targets using sonar sensors in real underwater environments.
This work focuses on 3D mapping and localization, aiming to enhance the mission performance of field robots through robust robotic perception. This research includes deep-learning-based scene recognition and precise 3D model reconstruction using low-quality sensors, large area mapping using multi-modal sensors like sonar, LiDAR, and optical cameras, and the development of a graph-based SLAM framework for field robots. The work was applied to underwater robots, achieving quantitative objectives through engineering basin experiments and demonstrating its effectiveness in real-world field environments.
This work addresses the challenge of two robots passing each other in narrow passages. The key technologies include 2D LiDAR-based mapping, potential-energy-based real-time path planning, and IoT-based local communication between robots. Through local communication, robots can compare their paths and priorities when encountering each other, allowing one robot to yield to the other. This project, in collaboration with the University of Tokyo, was applied to two mobile robots in 2024.