| 中村 維冴 | M, 2回目発表 | ロボットラーニング | 松原 崇充 | 和田 隆広 | 柴田 一騎 | 鶴峯 義久 | 佐々木 光 |
|
title: GCRL: Goal-Augmented Contrastive Reinforcement Learning for Long-Horizon Tasks
abstract: Contrastive Reinforcement Learning (CRL) is a goal-conditioned reinforcement learning method that estimates the value function through contrastive learning, achieving high sample efficiency in learning robotic tasks that were challenging for traditional RL. In CRL, various goals are sampled from experience data based on the current policy, and the goal-conditioned value function is estimated using contrastive learning. However, due to the inherent difficulty of sampling distant goal, value estimation becomes challenging, leading the problem of learning long-horizon robotic tasks. To mitigate this issue, we propose Goal-Augmented Contrastive Reinforcement Learning (GCRL). We firstly introduce Goal Augmentation to artificially increase the size of a dataset by applying various transformations to the original data. Secondly, we introduce a factorized value learning scheme from the perspective of Mutual Information, to prevent unreasonably high estimates of distant goal values that are inherently unreachable. This enables accurate value estimation even from goal-augmented samples. Through simulation analysis and comparative experiments across multiple tasks, we confirmed that proposed GCRL framework is indeed effective and that long-horizon robotic tasks, which were difficult to learn with conventional CRL, can be successfully learned by using our method. language of the presentation: Japanese | |||||||
| 矢野 嘉希 | M, 2回目発表 | ロボットラーニング | 松原 崇充 | 和田 隆広 | 柴田 一騎 | 鶴峯 義久 | 佐々木 光 |
|
title: Latency-aware Coach-Player Multi-Agent Reinforcement learning
abstract: Multi-robot control systems ideally combine centralized coordination signals from a fully observable server with autonomous decision-making by robots based on local observations. However, latency due to server-side inference can cause inconsistencies in the autonomous robots' observations, making task completion difficult. This research proposes a reinforcement learning framework that utilizes variables explaining latency trends available at runtime to acquire policy adaptable to latency fluctuations. Additionally, as a future prospect of this research, we propose a framework for achieving task completion through language instructions. language of the presentation: Japanese | |||||||
| 鉢峰 拓海 | D, 中間発表 | ロボットラーニング | 松原 崇充 | 和田 隆広 | 柴田 一騎 | 鶴峯 義久 | 佐々木 光 |
|
title: Automation of Object Shape Manipulation Tasks with Removal Process by Cutting Surface Model
abstract: Object-shape manipulation is widely applied in industry and our daily life. Removal process is a one of the shaping method that gradually removing unnecessary material from the base stock. Object-shape transition models are essential to achieve automation by robots; however, learning such a complex model that depends on process conditions (e.g., material type, robot posture, etc.) is challenging because it requires a significant amount of data, and the irreversible nature of the removal process makes data collection expensive. Above in this mind, this study proposes a scalable automation framework for object shaping by removal process. Therefore, we focus on the removal process is accomplished by local surface contact between the tool surface and the object. In other words, by interpreting that the shape transition is accomplished by the cutting surface, the shape transition can be represented by a geometric shape transition model with a cutting surface. This model can be commonly applied to a wide variety of removal process because the transitions are geometrically computable and require no learning. We applied the cutting surface model to two types of object shape manipulation tasks and evaluated the effectiveness of the proposed approach. language of the presentation: Japanese   | |||||||
| 福島 和希 | D, 中間発表 | ソフトウェア工学 | 松本 健一 | 飯田 元 | 石尾 隆 | Raula Gaikovina Kula | 嶋利 一真 |
|
title: Study on Instructor Support in Assignment-based Programming Courses Based on Source Code Similarity Analysis
abstract: In programming education within higher education, assignment-based programming courses, which are commonly used, present several challenges. For instance, there may be a disconnect between the skills that should be cultivated according to the course plan and those actually addressed, the difficulty of the assigned tasks may not be appropriate, and the workload related to grading and feedback can be overwhelming. This study focuses on these issues, categorizing them into two phases: before and after the implementation of the course. We conducted research on a method to assist instructors by utilizing source code similarity analysis techniques. language of the presentation: Japanese 発表題目: ソースコードの類似性分析に基づくプログラミング授業の講師支援に関する研究 発表概要: 高等教育のプログラミング教育においてよく用いられる課題ベースのプログラミング授業には様々な課題点が存在する．例えば授業計画において養成されるべき力との乖離があったり，提示される課題の難易度が適切でなかったり，評価やフィードバックについての作業負荷が大きかったりといった点である．これらの課題点について，大きく授業の実施前後で分け，ソースコードの類似度分析の手法を活用することにより，講師の授業支援を行う方法についての研究を実施した． | |||||||