| 研究実績の概要 |
During the current fiscal year, our project has concentrated on Aims 2 and 3 of the research proposal, advancing learning strategies for manipulation tasks. We have successfully developed simulation environments specifically tailored for the cable-driven gripper introduced last year. These environments facilitate interactions with deformable objects and the assessment of reinforcement learning strategies for complex manipulation tasks, such as applying tension or executing folds. Building upon the previous year’s work on the cable-driven gripper, we engineered a three-fingered underactuated robotic hand. This innovative hand features compliant joints and cable-driven actuation, with each finger being actuated by a single motor that manages both flexion and abduction movements. The hand boasts a total of four actuated degrees of freedom (DOFs). Significant progress has also been made in developing the control mechanisms and interaction capabilities of the new hand gripper. We created a simulation environment to train in-hand manipulation of small objects, implementing and validating several reinforcement learning algorithms for this purpose. In particular, Aim 3, the transition to real-world applications, is ongoing and constitutes the focus of our future work. Techniques such as contrastive learning, image-to-image translation, and transfer learning are proving instrumental in rendering our simulation environments more representative of actual scenarios.
|
