Hi, I’m Dinithi. PhD candidate at the National University of Singapore, in the Augmented Human Lab, working with Prof. Suranga Nanayakkara.

Wearables that sense the person, not just the scene.

I build wearable AI systems that use egocentric audio, video, and physiological signals to understand a wearer's context and what they can take in at any given moment. By adapting when, how, and whether AI should intervene, I aim to make wearables more intuitive, and seamlessly integrated into everyday life.

Try the idea behind HeadRoom Which of your senses has room for a notification?

Eyes —

Ears —

Where are you? Quiet desk Busy café In a conversation

Scroll or move your pointer to make your eyes busier.

A browser sketch, not the model. Here your scrolling and pointer movement stand in for visual load and the scene sets audio load. HeadRoom itself estimates availability from how unpredictable the next moment of egocentric video and audio is.

UIST 2026 HeadRoom accepted as a paper and a demo; I’m presenting it on 5 November at 9 am in Detroit. I’m also presenting my thesis on user-aware adaptive assistive wearables at the Doctoral Consortium.
Distinguished Poster Award Research Summit for Egocentric Intelligence 2026, organised by Meta
3rd place Emotional Mimicry Intensity Estimation Challenge, Affective Behavior Analysis in-the-wild workshop at CVPR 2026

Looking for research internships in 2027. 12–24 weeks, in applied AI, HCI, XR and wearables.

Email me

My vision

Wearables that adapt to the person wearing them. Most assistants react to the scene: what the camera sees and the microphone hears. I want them to also read the wearer, their attention, load and state, and change when and how they help.

Reading the person

Ongoing Sensory Spotlight Forecasts when the wearer’s attention will shift, where it will land, and what they will reach for next. 3rd place, CVPR 2026 workshop challenge Emotion mimicry intensity A lightweight multimodal model that estimates how strongly someone mirrors an emotion. Case study Looks Fine to Me Behavioral and physiological signs that a person has stopped really checking an AI agent’s work.

Adapting the assistance

UIST 2026 HeadRoom Routes each notification away from the sense that is overloaded: to the display, or to audio. Case study, CHI 2026 Interrupting Autopilot Sensory interruptions that help people catch and correct a habit in the moment.

I build the models and the systems, then test them with people. Grounded in our systematic review of 63 papers on how wearables sense, adapt and intervene.

Projects

UIST 2026, paper and demo

HeadRoom

A lightweight, edge-deployable pipeline for adaptive notification routing

HeadRoom predicts the next moment of egocentric video and audio. When a stream is hard to predict, that sense is probably busy. It routes each notification to the channel with more room. In a controlled study with 25 participants, routing to the more available channel reduced response time under high perceptual load compared with routing to the less available one.

Runs in real time on wearable and mobile edge hardware
Distinguished Poster Award, Meta’s Research Summit for Egocentric Intelligence 2026

Project page

HeadRoom pipeline: egocentric video and audio features are predicted one step ahead; the prediction errors for each channel feed a routing module that sends the notification to the glasses display or to audio. — Prediction error on each stream stands in for how busy that sense is.

Ongoing, unpublished

Sensory Spotlight

Anticipating shifts in attention for proactive assistance

Proactive assistance depends on timing. Support that arrives after attention has moved is easily missed, and running every sensor continuously is costly for a device worn all day. Sensory Spotlight investigates how lightweight wearable signals, such as eye and head movement, can anticipate when a wearer’s attention is about to shift and where it will land, so that assistance, and the more demanding sensors behind it, are engaged only when they are needed.

Under review

User-aware adaptive assistive wearables

A systematic review of 63 papers

How do adaptive wearables sense user state, decide when to act, and choose how to intervene? We propose a taxonomy of sensing modalities, adaptation triggers and intervention strategies, and set out the open design challenges for the next generation of assistive wearables.

Taxonomy of adaptive assistive wearables covering sensing, adaptation triggers and interventions.

Under review

Looks Fine to Me

Understanding oversight in agentic interactions

As AI agents take over execution, the human job becomes oversight: noticing when an output needs a closer look. In a within-subject study with 36 participants, people completed Python tasks with a coding agent that returned correct code, a silent defect or a visible failure, while we recorded egocentric video, gaze, physiological signals and interaction logs.

Participants caught visible failures readily but often missed silent defects while feeling highly confident. They ran generic checks, did not gather diagnostic evidence, or misread the evidence they had. Feeling vigilant and being vigilant are not the same thing.

Read the preprint (arXiv)

Study setup: a participant works with a coding agent while egocentric, physiological and interaction data are recorded. The agent returns code with no defect, a blatant defect or a silent defect, and oversight is traced from general evaluation to task-relevant evaluation, defect detection and resolution.

CHI 2025 Late-Breaking Work

VRSense

An explainable system to help mitigate cybersickness in VR games

A plug-and-play tool that helps VR developers assess cybersickness. It uses interpretable features instead of a black box, so developers see which design choices and interactions are making players sick.

Read the paper

CVPR 2022 and ECCV 2022 workshop

Privacy for 3D objects in mixed reality

Transforming and regenerating point clouds on device

A 3D–2D correspondence technique for point clouds and a 3D vision method that lets users add, delete or modify parts of an object before it is regenerated on the other end. Evaluated against simulated privacy attacks and run on device.

CrossPoint (CVPR) 3DLatNav (arXiv)

Mixed reality 3D object privacy pipeline.

Papers

Full list on Google Scholar.

2026HeadRoom: lightweight, edge-deployable pipeline for adaptive notification routingUIST 2026 (paper and demo)
2026Interrupting Autopilot: evaluating sensory interventions for momentary self-regulationCHI 2026 Late-Breaking Work
2026Two-stage multimodal framework for emotion mimicry intensity predictionABAW Workshop, CVPR 2026
2025VRSense: an explainable system to help mitigate cybersickness in VR gamesCHI 2025 Late-Breaking Work
20223DLatNav: navigating generative latent spaces for semantic-aware 3D object manipulationECCV 2022 workshop
2022CrossPoint: self-supervised cross-modal contrastive learning for 3D point cloud understandingCVPR 2022

Experience

2023 – now

PhD candidate, National University of Singapore

Augmented Human Lab, School of Computing. Passed the qualifying exam in September 2024.
2024 – 2025
Research collaboration, Meta Reality Labs × Augmented Human Lab
Multimodal XR user studies and predictive modelling of cybersickness.
- Designed and ran a VR gameplay study with 150 participants on motion sickness during active play.
- Collected and synchronised eye tracking, head motion and physiological signals.
- Built models that predict discomfort and cybersickness onset in real time.
- Built pipelines that align heterogeneous sensor streams at scale.
- Turned model outputs into actionable insights for VR game evaluation.
2023
Data analytics consultant, LIRNEasia
Geospatial ML and longitudinal datasets for policy research.
- Classified built-up regions from satellite imagery for a national urbanisation index.
- Automated spatial analysis with QGIS and Google Earth Engine.
- Built longitudinal datasets on household energy consumption.
- Worked with policy researchers to turn findings into evidence for urban planning.
2022 – 2023
Data engineer, Axiata Digital Labs
Production ML pipelines across telecom domains on AWS and GCP.
- Engineered ML/DL deployment pipelines for the company’s AI Factory platform.
- Maintained data ingestion and model integration across telecom business domains.
- Built and deployed an end-to-end customer churn prediction model for enterprise clients.
- Bridged research prototypes and product with engineering teams.
2017 – 2022

BSc Electronic & Telecommunication Engineering, University of Moratuwa

First Class Honours. Final-year project on privacy for 3D objects in mixed reality.

News

Sep 2026Accepted to the UIST 2026 Doctoral Consortium with my thesis on user-aware adaptive assistive wearables.
Jul 2026HeadRoom won the Distinguished Poster Award at the Research Summit for Egocentric Intelligence 2026, organised by Meta.
Jul 2026HeadRoom accepted to UIST 2026.
Apr 2026Presented our paper at the Affective Behavior Analysis in-the-wild workshop, CVPR 2026 in Denver.
Mar 20263rd place in the Emotional Mimicry Intensity Estimation Challenge at the Affective Behavior Analysis in-the-wild workshop, CVPR 2026.
Feb 2026“Interrupting Autopilot” accepted to CHI 2026 Late-Breaking Work.
Jan 2026Submitted our systematic review of user-aware adaptive assistive wearables.
Aug 2025Talk on adaptive cognitive augmentation at Meta Reality Labs, Redmond.
Feb 2025VRSense accepted to CHI 2025 Late-Breaking Work.
Sep 2024Started working on the AHLab × Meta Reality Labs research collaboration.
Sep 2024Passed my qualifying exam and became a PhD candidate.
Aug 2023Started my PhD at NUS.
Jan 2023Started data analytics consulting at LIRNEasia.
Oct 20223DLatNav presented at an ECCV 2022 workshop.
Aug 2022Graduated with First Class Honours from the University of Moratuwa.
Jun 2022CrossPoint presented as a full paper at CVPR 2022.
May 2022Joined Axiata Digital Labs as a data engineer.