PLAY · IMAGINE · APPLY
Program of Research
Norm-Competent AI Agents for Human Teams, Companions, and Training
I design experiments, cognitive architectures, and deployed training systems that make AI agents norm-competent — able to follow, reason about, and legibly deviate from the tacit rules that govern what a good teammate, teacher, or companion may, must, and must not do. My goal is for human trust in AI agents to stay calibrated to what those agents actually deserve.
Research Contribution Areas
1. The Problem and Three Research Strands
AI agents are moving into the settings people care most about: a robot teammate in a safety-critical procedure; an AI tutor shaping how a medical resident learns to reason; a conversational agent that becomes, for many users, a daily companion. Every one of these settings runs on norms — yet systems trained above all to be engaging and agreeable flatter learners when they should challenge them, keep conversations going when they should end them, and cannot explain when or why a rule should bend. The result is miscalibrated trust in exactly the settings where miscalibration is most costly.
My research program addresses this through a build–measure–deploy cycle: findings from controlled human-subjects experiments become architectural requirements; architectures become deployed systems; deployments generate the next round of questions. This methodology spans three connected strands.
Research Strand 1
Trust & Norm Violation in Human-AI Teams
How does an AI agent's norm behavior — adherence, violation, and disclosure — shape the trust humans place in it, and can a committed disclosure policy outperform ad hoc explanation?
Experimental studies using within-subjects designs crossing norm adherence with violation outcome test the higher-good hypothesis — that a robot's norm violation can preserve or increase trust when legibly in service of the team. Further studies import Bayesian persuasion and calibrated disclosure (building on information-design theory) to ask whether a robot's committed disclosure policy outperforms ad hoc explanation on trust calibration.
Research Strand 2
Social Robots, Affect, and Norm-Competent Architectures
How do social robots read, model, and respond to human affect and social norms — and how can that reasoning be made explicit, auditable, and verifiable rather than emergent and opaque?
Social robots don't just act — they are perceived as social agents, capable of warmth, attentiveness, and moral failure. This strand takes seriously the full arc of human-robot interaction: affect recognition, relationship formation, and the normative fabric that governs what a good companion may, must, and must not do. It draws on HRI, affective computing, and deontic social theory to ask how robots can participate in human social life without misrepresenting their nature or eroding the trust they depend on.
The technical anchor is the Klara Cognitive Simulator — built around Kazuo Ishiguro's Klara and the Sun, in which an Artificial Friend navigates the norms of companionship, loyalty, and sacrifice with no guarantee of reciprocity. Literary fiction encodes exactly the socially rich, affect-laden, norm-dense situations that benchmark suites lack: Klara must decide when to tell the truth, when silence is kinder, and when a rule should bend for someone she loves. These scenarios become the evaluation corpus for norm-aware agent pipelines, tested against a hybrid architecture where a symbolic deontic layer holds veto authority over LLM generation — making norm reasoning an explicit, auditable object rather than a statistical tendency absorbed from training data.
Working with M. Scheutz in the DIARC lineage, I develop and evaluate these pipelines against deontically specified social scenarios. Released openly, Klara is designed as a shared community instrument for the field.
Research Strand 3
AI-Enabled Serious Games
How to design "smart-enough" serious games?
Serious games — environments where play is inseparable from consequential action — offer a uniquely tractable design space for applying AI: the rules of the game are explicit, player behavior is observable, and the gap between intended and actual AI behavior is measurable. This strand develops a principled framework for designing and validating serious games in the era of AI, grounded in a distinction the field had been conflating: instructional intelligence (the game's capacity to infer player state) and instructional adaptivity (its capacity to respond to it) are separable subsystems with distinct failure modes.
A system can be intelligent without being adaptive — and, more dangerously, adaptive without being intelligent: fluent real-time responsiveness driven by an inaccurate model of the player. This reframes the central design question from "how sophisticated is the AI?" to "is the coupling between inference and action validated for this game's purpose?"
Research Strand 4
Precision Medical Education & AI-Based Systems
Can AI-based systems deliver the full precision medical education cycle — and what couplings between intelligence and adaptivity does that require?
Precision medical education (PME) calls for instruction that continuously adapts to each learner's evolving profile through a cyclical loop of observation, inference, intervention, and measurement. AI-based systems can close this loop — but only when intelligence (the quality and variety of what a system can produce) and adaptivity (the degree to which that production tracks the individual learner's state) are treated as distinct, separable axes that must be aligned. The intelligence layer is advancing rapidly; the adaptivity layer — where pedagogical expertise is decisive — is where the field's next decade of work belongs.
A concrete direction in this strand is the design of AI-based platforms for automated, asynchronous assessment of clinical reasoning in graduate medical education. These systems embed assessment in authentic clinical workflows — such as medication reconciliation at discharge — and require trainees to document structured reasoning (decision → rationale → cited evidence) for every clinical choice. A hybrid assessment architecture pairs deterministic error-taxonomy scoring against educator-authored gold standards with LLM-based argumentation mining, which constructs credibility-weighted argument graphs and maps reasoning quality to ACGME milestone sub-competencies. The goal is immediate, granular feedback on both what a trainee decided and why they reasoned as they did — a distinction no checklist score alone can make.
2. Current Research at Tufts HRI Laboratory
A foundational question in human-robot and human-AI interaction is not whether agents should follow rules — but what happens to human trust when they don't. My research program investigates the conditions under which an AI or robot agent's norm violation can preserve or even increase trust, rather than erode it. This requires separating the normative act of deviation from its downstream outcome — a distinction most prior work has conflated.
The current within-subjects study crosses norm adherence with violation outcome, holding communication style constant, to test the higher-good hypothesis: that a robot's principled deviation — one legibly in service of the team goal — can function as a trust-warranting signal rather than a trust-breaking one. This study is grounded in deontic social norms theory and human-robot teaming literature, and uses validated trust instruments (Schaefer, Godspeed, NASA-TLX) alongside objective task metrics.
Two follow-up studies extend the arc. The first disentangles outcome bias from norm-based judgment through an outcome-valence manipulation — asking whether participants are responding to the success of the deviation or the principled nature of the deviation itself. The second imports formal information design — Bayesian persuasion and calibrated disclosure — to test whether a robot's committed disclosure policy about its own deviations outperforms ad hoc explanation on trust calibration. To our knowledge, this would be the first experimental test of information-design predictions in the context of human-robot norm conflict.
Social robots occupy a unique design space: they are perceived as social agents with affect, intent, and moral standing — yet built from systems that have none. This project takes that gap seriously. Rather than treating affect and norms as separate engineering problems, it asks how a robot can legibly model and respond to human emotion while navigating the normative structure of a relationship. The question is not only whether a robot follows rules, but whether it does so in a way that feels earned.
The design challenge is grounded in Kazuo Ishiguro's Klara and the Sun — a novel whose protagonist, an Artificial Friend named Klara, must reason about loyalty, sacrifice, and the limits of what she is permitted to do for someone she loves. Klara makes an ideal specification target: she experiences affect without being certain of its meaning, follows norms without always understanding their origin, and must decide when to bend a rule for a higher good. The Klara Cognitive Simulator operationalizes these scenarios as a rigorous evaluation corpus for norm-aware agent pipelines.
Working with Matthias Scheutz in the DIARC tradition, I aim to develop hybrid symbolic–LLM architectures where a deontic layer (obligations, prohibitions, permissions) holds veto authority over generation — making norm reasoning an explicit, auditable object rather than a statistical tendency absorbed from training data. Released openly, Klara is intended as a shared community instrument for benchmarking social robot and AI companion behavior in norm-dense, affectively rich interaction.
4. Research Publications by Area
The following prior research programs provide the empirical and technical foundation for the current agenda. Each thread is now being extended into the current research agenda at Tufts.
Dietra is a reinforcement learning agent for individualized diabetes self-management — modeling meal timing, activity patterns, and glycemic trends to generate adaptive, personalized coaching.
Dietra: RL-Based Diabetes Coaching
Full clinical system (access-controlled) · dietra.piatripathi.ca
Reinforcement Learning-Driven Nutrition Coaching
Tripathi, P., Gupta, N., & Kahol, K.
CASCON 2025, Toronto, CA, 2025.
View Publication →Applied work in competency-based education and AI-enabled serious games for health professions, grounded in the instructional intelligence/adaptivity framework.
In Press
Intelligence and Adaptivity Are Not the Same: Toward a Design Framework for Adaptive Learning Environments
Tripathi, P.
Advances in Global Applied AI, Springer. Preprint: arXiv:2605.21962, 2026.
View Publication →Accepted
LLM Odyssey: Design of a Game-Based Platform for Teaching LLM Engineering Concepts
Tripathi, P.
Frontiers in Education (FIE) 2026, Work-in-Progress. Camera-ready submitted., 2026.
Automated gesture segmentation and motion recognition using coupled HMMs, providing the perceptual substrate for agents that interpret physical human communication. This work directly motivates current architectures for reading embodied social signals in human-robot interaction.
Automated Gesture Segmentation From Dance Sequences
Tripathi, P., et al.
IEEE Automatic Face and Gesture Recognition, 2004.
Documenting motion sequences with a personalized annotation system
Kahol, K., Tripathi, P., & Panchanathan, S.
IEEE MultiMedia, vol. 13, no. 1, 2006.
View Publication →Context-sensitive haptic rendering, visio-haptic attention modeling, and rehabilitation system design established AI systems that communicate through touch — directly motivating agents that communicate intent physically in human-robot interaction.
Modeling context in haptic perception, rendering, and visualization
Tripathi, P., et al.
ACM TOMCCAP, 2006.
A Model for Visio-Haptic Attention for Efficient Resource Allocation in Multimodal Environments
Tripathi, P., Kahol, K., et al.
Foundations of Augmented Cognition, Springer, 2007.
View Publication →Rehabilitation of patients with hemispatial neglect using visual-haptic feedback
Tripathi, P., et al.
HCI International, 2005.
Under Revision
When I Say... Pseudohaptics
Tripathi, P., & Kapralos, B.
Medical Education. Major revision invited. Revised manuscript due October 2026., 2026.
PhD research on experience-sampling and sociometric modeling of creative teams produced behavioral measurement frameworks — now being generalized to human-AI collaborative dynamics including synchrony quantification and trust modeling in human-robot teams.
Predicting Creativity in the Wild: Experience Sample and Sociometric Modeling of Teams
Tripathi, P. & Burleson, W.
CSCW, ACM, 2012.
View Publication →Mining Creativity Research to Inform Design Rationale in Open Source Communities
Burleson, W. & Tripathi, P.
Human Technology Journal, 2011.
View Publication →