| 角川 勇貴 | D, 中間発表 | ロボットラーニング | 松原 崇充, 和田 隆広, 鶴峯 義久 |
|
title: Deep Reinforcement Learning Framework for Edge Robot Control
abstract: Recently, edge robots with built-in computers are expected to take autonomous actions. Deep reinforcement learning (DRL), which can obtain control policies from trial-and-error action samples, is expected to be applied to these robots. However, the framework of obtaining policies from DRL can be divided into two major processes: Policy learning and utilization. However, typical frameworks are difficult to apply because they do not consider the limitations of activity time and calculation speed that are unique to edge robots. Specifically, the limited calculation speed of the DRL algorithm makes real-time utilization difficult. Policy learning is difficult because of the difficulty of collecting training samples due to the limited activity time. Thus, in this study, we addressed each of the following. 1) Limitation of calculation speed: Accelerating the calculation speed proposing DRL theory utilizing FPGA and binarized neural networks. 2) Activity time limitation: Reduction of training sample cost by obtaining the samples by simulation and proposing stable DRL with simulation. The effectiveness of the proposed method was verified in real-robot tasks: a real-time visual-servo task and a ball-breaking task with a vast transition pattern. language of the presentation: Japanese | |||
| 畠中 渉 | D, 中間発表 | ロボットラーニング | 松原 崇充, 和田 隆広, 鶴峯 義久 |
|
title: Learning semi-autonomous system to handle uncertainty in workflow progress and operator's decision
Demand for workplace automation through AI and robots is growing due to labor shortages. The semi-autonomous system involves cooperative work between humans and robots and is expected to be implemented earlier than the fully autonomous system. However, it is challenging to control the robot consistently satisfying the ordered workflow procedure because of the uncertainty involved in the human operator's decision. To address this problem, we propose a learning algorithm that enables the robot to branch its behavior in response to the degree of ambiguity in the operator's decision. We manage the workflow process by Linear Temporal Logic (LTL), a simple and expressive formal language, and capture the uncertainty in the workflow progress as a state by defining a belief LTL by the ambiguity in the operator's decisions. The belief LTL is encoded by a graph neural network, and the policy is optimized by reinforcement learning to follow a given workflow without violating it. Furthermore, our method can reduce unnecessary operations and the workload on the operator by optimizing the timing of queries to the operator simultaneously. We conduct experiments in a 3D simulation environment assuming an inspection task by a robotic arm and show that learned policy can behave appropriately depending on the uncertainty of the workflow progress. language of the presentation: Japanese | |||
| 山之口 智也 | D, 中間発表 | ロボットラーニング | 松原 崇充, 和田 隆広, 鶴峯 義久 |
|
title: Model learning for sim-to-real transfer enabling selection of task-relevant domain information
abstract: Model predictive control is widely used in robot control as a method that is robust to modeling errors and can be applied to various tasks. However, data collection cost in the real world is very high, making it difficult to train model in the real world. To overcome this problem, sim-to-real transfer approaches, in which the model is trained in simulation and then transferred to the real world, have been explored in recent years. While such approaches have been validated in several studies, task-oriented sim-to-real transfer has not been achieved due to the lack of selection of the domain information necessary and unnecessary for the task to perform the sim-to-real transfer. In this research, we propose a model learning algorithm that enables the selection of task-relevant domain information for sim-to-real transfer. We empirically verified the effectiveness of the proposed method by applying it to two domain-independent tasks, including a valve rotation task and a block mating task. language of the presentation: English | |||
| 前田 雄大 | D, 中間発表 | 計算システムズ生物学 | 金谷 重彦, 安本 慶一, 小野 直亮, MD.ALTAF-UL-AMIN, 黄 銘 |
| YANG ZIWEI | M, 2回目発表 | 計算システムズ生物学 | 金谷 重彦, 安本 慶一, 小野 直亮, MD.ALTAF-UL-AMIN, 黄 銘 |
|
title: An Unsupervised Learning Framework for Identifying Cancer Subtypes
abstract: Cancer is one of the deadliest diseases worldwide. Accurate diagnosis and classification of cancer subtypes are indispensable for effective clinical treatment. Promising results on automatic cancer subtyping systems have been published recently with the emergence of various deep learning methods. However, such automatic systems often overfit the data due to the high dimensionality and scarcity. In this research, we propose to investigate automatic subtyping from an unsupervised learning perspective by directly constructing the underlying data distribution itself, hence sufficient data can be generated to alleviate the issue of overfitting. Specifically, we bypass the strong Gaussianity assumption that typically exists but fails in the unsupervised learning subtyping literature due to small-sized samples by vector quantization. In this manner, the proposed method is able to better capture the latent space features and model the cancer subtype manifestation on a molecular basis. language of the presentation: English | |||