コロキアムB発表

日時: 07月29日 (Tue) 1限目(9:20-10:50)


会場: L2

司会: 松井智一
MUHAMMAD AKMAL BIN MOHAMMED ZAFFIR D, 中間発表 ヒューマンロボティクス 和田 隆広, 清川 清, 劉 海龍, 織田 泰彰, 本司 澄空
title: Towards a Proactive Robot-to-Human Object Handover Framework using Forearm-Wearable Device
abstract: This study proposes a proactive robot-to-human (R2H) object handover framework using a forearm-wearable vibrotactile device that guides the human hand toward the object transfer point (OTP). The aim is to improve the speed and accuracy of human reaching movements during handovers. To enable precise human reaching response to the presented vibrotactile cues (human response), we first investigated the effects of vibrotactile signal parameters—amplitude, duration of stimulus (DoS), and inter-stimulus interval (ISI)—on both perceptual and physical responses. Results show that tuning these parameters does not significantly enhance responses precision, highlighting the variability in human reaction. To address this, we propose a modeling approach implementing Heteroscedastic Gaussian Process Regression (H-GPR) in a recursive-sparse manner, which captures the stochastic and location-dependent uncertainty in human response. Based on this model, we propose a system that adaptively selects the optimal vibrotactile directional cue to minimize uncertainty. The proposed system key idea will be presented, demonstrating its potential in facilitating smoother and more intuitive R2H interactions.
language of the presentation: English
 
山本 愛奈 M, 2回目発表 ヒューマンロボティクス 和田 隆広, 清川 清, 劉 海龍, 織田 泰彰, 本司 澄空

title: Verification of the Robustness of Model Parameters 

Representing Individual Progression of Motion Sickness Symptoms

abstract: In recent years, with the spread of self-driving cars, it is believed that motion sickness will increase due to the diversification of non-driving activities. As a result, research on modeling motion sickness has been actively conducted in order to quantitatively evaluate and predict it, and to develop countermeasures. Previous studies have used models that predict the progression of individual MISC scores to examine how well future symptom progression can be predicted from limited past data, showing that utilizing symptom history reduces prediction error. However, the generalizability of model parameters across different frequency conditions has not been sufficiently investigated. Therefore, this study examines whether parameters identified under specific frequency conditions can be applied to other frequency conditions.

language of the presentation: Japanese

発表題目: 動揺病症状の個人進行を表現する計算モデルのパラメータのロバスト性検討

発表概要: 近年,自動運転車の普及に伴う運転外活動の多様化により,動揺病の増加が懸念されている.そこで,動揺病を定量的に評価・予測し,対策を講じるために,動揺病のモデル化に関する研究が活発に行われている.先行研究では,個々のMISCの進行を予測するモデルを用い,限られた過去データから将来の症状進行をどの程度予測できるかを検証し,症状履歴の活用によって予測誤差が低減することが示された.しかし,異なる周波数条件に対するパラメータの汎用性については検証が不十分である.そこで本研究では,特定の周波数条件で同定されたパラメータが他の周波数条件にも適用可能かどうかを検証する. 

 
井出 優哉 M, 2回目発表 ヒューマンロボティクス 和田 隆広, 清川 清, 劉 海龍, 織田 泰彰, 本司 澄空
title: Reducing Motion Sickness in Passengers of Autonomous Personal Mobility Vehicles by Presenting a Driving Path
abstract: Autonomous personal mobility vehicles (APMVs) are small mobility devices designed for individual automated transportation in shared spaces. In such environments, frequent pedestrian avoidance maneuvers may cause rapid steering adjustments and passive postural responses from passengers, thereby increasing the risk of motion sickness. This study investigated the effects of providing path information on 16 passengers' head movement behavior and motion sickness while riding an APMV. Through a controlled experiment comparing manual driving (MD), autonomous driving without path information (AD w/o path), and autonomous driving with path information (AD w/ path), we found that providing path cues significantly reduced MISC scores and delayed the onset of motion sickness symptoms. In addition, participants were more likely to proactively align their head movements with the direction of vehicle rotation in both MD and AD w/ path conditions. Although a small correlation was observed between the delay in yaw rotation of the passenger's head relative to the vehicle and the occurrence of motion sickness, the underlying physiological mechanism remains to be elucidated.
language of the presentation: Japanese
発表題目: 経路提示による自動運転パーソナルモビリティビークル乗客の動揺病低減手法の検討
発表概要: 近未来,自動運転パーソナルモビリティビークル(APMV)が歩行者の多い共有空間で使用されると考えられる.その際,歩行者との遭遇が頻繁に起こり,APMVの動きに対して受動的な動きをしてしまい,乗客の動揺病の進行が早まる可能性がある.そこで,本研究ではAPMVの将来経路を16名乗客に提示することによる,乗客の補償行動の促進および動揺病の進行防止の効果について調査した. 手動運転(MD), 経路情報なしの自動運転(AD w/o path), および経路情報付きの自動運転(AD w/ path)を比較した, 乗客16名による実験を実施した結果, 動揺病の主観的な評価項目であるMISCの値が経路情報に比べ, 有意に減少することが確認された. さらに, MDとAD w/pathの両条件で車両のYaw方向の回転に乗客が頭部の動きを一致させる傾向がみられた. 乗客の頭部の動きと動揺病の発症には小さな相関が確認されたが, 直接的な原因となったかは依然として不明である.
 

日時: 07月29日 (Tue) 1限目(9:20-10:50)


会場: L3

司会: Faisal Mehmood
SUN LEI D, 中間発表 数理情報学(コミュニケーション学) 池田 和司☆, 岩田 具治, 田中 佑典, 久保 孝富

title: Meta-learning from pre-trained models and unlabeled data

abstract: Meta-learning without labeled data is crucial for real-world applications, as labeled datasets are often costly to acquire or restricted due to privacy concerns. Data-Free Meta-Learning (DFML) addresses this issue by enabling meta-learning without access to training data, by leveraging pre-trained models. However, existing DFML methods face two primary practical challenges: high computational costs caused by model inversion-based data recovery, and the underutilization of unlabeled data available in real-world scenarios. We propose a meta-learning method that learns from pre-trained models and a small unlabeled dataset without relying on data recovery. Our method generates meta-training tasks by assigning soft labels to unlabeled data using pre-trained models. Since the generated meta-training tasks vary widely in quality, we introduce a task-weighting mechanism based on task confidence and class distribution balance for effective meta-learning. Extensive experiments demonstrate that our approach significantly reduces the computational cost and enhances generalization, achieving up to 111$\times$ speed-up and 12.7\%-58.2\% few-shot classification performance improvement compared to the state-of-the-art DFML method. Results also show that our method generalizes effectively across different domains and model architectures.

language of the presentation: Japanese


 
植原 真人 D, 中間発表 数理情報学(計算神経科学) 池田 和司☆, 川鍋 一晃, 杉本 徳和, 田中 沙織
title: Current source reconstruction method in the frequency domain
abstract: Current source reconstruction is a method for estimating cortical current sources based on measured electroencephalogram (EEG) and magnetoencephalogram (MEG) signals. Given that the number of cortical current sources is usually greater than the number of observed sensors. To address this challenge, various methodologies have been proposed. most of these methods are applying filters on time series signals; however, they are not considering the frequency components of the reconstructed signals. Therefore, we propose a method for current source reconstruction in the frequency domain to reconstruct frequency characteristics of brain activity. Utilizing an estimation based on hierarchical variational Bayesian methods achieve to introduce a prior distribution for brain activity and sparse current source estimation. To assess the efficacy of the proposed method, simulations were conducted and the characteristics of the method were investigated. The simulation results shows that the proposed method accurately reconstructed time series data and estimated both the position of current sources and the frequency characteristics.
language of the presentation: Japanese
 
CHU SHIWEN M, 2回目発表 数理情報学(計算神経科学) 池田 和司☆, 川鍋 一晃, 杉本 徳和
title: Enhancing Cross-subject Emotion Classification Performance Using Resting-state EEG and Multimodal Features.
abstract: Recognizing human emotion from EEG signals is a significant challenge, primarily due to the Inter-subject variability across different individuals. To address this, our research introduces a novel transfer learning framework designed to improve the cross-subject generalization of emotion recognition models.Our key innovation is a subject adaptation strategy that personalizes a general model for a new user by analyzing a short, easily-recorded sample of their baseline resting-state EEG. This allows the model to adapt without requiring any new emotion-labeled data from the user. Furthermore, to enhance classification accuracy, our approach dynamically fuses the information from internal brain signals with external emotional cues from facial expressions using an attention-based mechanism.Experiments on public datasets demonstrate that our proposed method significantly enhances the model's ability to accurately classify emotions for new individuals compared to standard approaches.
language of the presentation: English