📚 Research Digest

Finding: This PhD research proposes reframing AI in physical activity tracking from prescriptive analytics to collaborative partnership, supporting users across the entire PA journey (goal setting, planning, tracking, reflection) while fostering user agency and self-efficacy.

Why it matters: Directly addresses human-AI collaboration, goal-setting behavior change, and motivation systems - core areas for understanding how AI can enhance rather than replace human decision-making and self-determination.

Method: Mixed methods approach investigating new interaction mechanisms and transparency models for human-AI collaborative systems.

Jargon: PA = Physical Activity; Self-efficacy = belief in one's ability to execute behaviors necessary to produce specific performance attainments.

Finding: A field study (N=167) comparing LLM-based chatbot vs. mobile app for smoking cessation found that chatbots excelled at empathy and social presence while apps were better for goal-setting and progress tracking, suggesting hybrid interfaces could leverage both strengths.

Why it matters: This directly informs HCI design decisions about when to use conversational vs. structured interfaces, particularly relevant for behavior change applications and understanding the tradeoffs between relational engagement and goal-oriented scaffolding.

Method: Comparative field study where participants chose between two interventions, used them for a week, then completed usability and effectiveness measures plus qualitative interviews.

Jargon: LLM = Large Language Model (AI system trained on text to generate human-like responses); persuasive interfaces = systems designed to influence user behavior.

Finding: A collaborative scientist-robot system was deployed in a planetary analog field mission where scientists could specify their knowledge, hypotheses, and preferences through a computer interface to guide autonomous robot data collection, enabling more efficient scientific exploration than traditional methods.

Why it matters: This demonstrates effective human-AI collaboration in scientific work, showing how interfaces can enable humans to communicate domain expertise to autonomous systems while leveraging AI's optimization capabilities.

Method: Field deployment study at White Sands National Park using a quadruped robot with three mission scientists who had different priors and objectives, measuring surface stiffness through ground-leg interactions.

Jargon: Planetary analog - Earth environments that simulate conditions on other planets for testing mission protocols.

Finding: Researchers developed laboratory tasks to measure psychological safety in human-robot teams, finding that participants performed better with human teammates than robot teammates across repeated trials.

Why it matters: This provides validated methods for studying team dynamics and psychological safety in human-AI collaboration, which is crucial for understanding how AI impacts workplace collaboration and team performance.

Method: 33 participants completed two different tasks repeatedly with both human and robot teammates in controlled laboratory conditions designed to simulate real operational environments.

Jargon: Psychological safety - team members' shared belief that they can take interpersonal risks without negative consequences; HRI - human-robot interaction.

Finding: Online learning environments weaken student self-efficacy compared to mixed-mode learning, but strategic interventions like synchronous touchpoints, scaffolded assignments, and trauma-informed teaching can restore student confidence and engagement in virtual classrooms.

Why it matters: Provides evidence-based strategies for designing online learning experiences that address motivation and goal achievement, directly applicable to teaching HCI, creativity, and behavioral science in digital environments.

Method: Theoretical framework drawing on Bandura's social cognitive theory with practical pedagogical recommendations based on contemporary research evidence.

Jargon: Self-efficacy = belief in one's ability to execute actions needed to achieve specific goals; Trauma-informed teaching = pedagogical approach that recognizes and responds to the impact of traumatic experiences on learning.

Finding: Two hybrid deep learning architectures combining audio spectrograms and video frames achieved 79-82% accuracy in classifying six emotions, with the EfficientNet-based model outperforming the ResNet18-ViT-LSTM approach on the CREMA-D dataset.

Why it matters: Improved emotion recognition systems directly enhance human-computer interaction by enabling more responsive and empathetic interfaces, and inform behavioral science research on emotional expression and recognition.

Method: Late fusion approach combining convolutional networks for feature extraction from audio spectrograms and video frames with LSTM networks for temporal pattern learning.

Jargon: CREMA-D: Crowdsourced Emotional Multimodal Actors Dataset; Late fusion: combining predictions from separate models rather than merging raw data; ViT: Vision Transformer architecture.

Finding: A hybrid system combining machine learning with Extended Kalman Filtering achieves real-time hand pose tracking using only 6-axis IMUs, improving joint angle accuracy by 40% over baseline methods while being robust to environmental interference.

Why it matters: This enables more natural and reliable hand-based interaction modalities for HCI applications, particularly in environments where vision-based tracking fails due to occlusion or lighting issues.

Method: Combines a learning-based pose estimation model trained on drift-insensitive features with an Extended Kalman Filter for noise reduction, evaluated on 12 hours of data across 23 tasks with 10 participants.

Jargon: IMU = Inertial Measurement Unit (sensors measuring motion and orientation); EKF = Extended Kalman Filter (statistical method for estimating system state from noisy measurements).