cover

DART: Robot-Assisted CBT Platform (NIH Research)

A web and cloud platform that lets a low-cost social robot guide a six-week CBT program at home. I lead the web side, and the system is described in our arXiv preprint.

2026-01-01

Next.jsTypescriptReactTailwind CSSLLMClaude Code
TeamWebAIHCIHuman-Centered AILLM ApplicationsReal-time SystemsIn Progress
A web and cloud platform that lets a small, low-cost robot guide students through six weeks of CBT exercises at home.

πŸ“„ Paper: A Deployable Architecture for Robot-Mediated Tasks (DART): Evaluation in Socially Assistive Robot-Guided Cognitive Behavioral Therapy Exercises (arXiv preprint, 2026). I'm one of the co-authors.

πŸ§‘β€πŸ’» Role: Web Team Lead at the USC Interaction Lab (Prof. Maja MatariΔ‡), on an NIH R01-funded project


🧠 Background

Our lab studies whether a robot named Blossom can help university students with anxiety keep up with their CBT homework. Blossom costs about $400 and runs on a Raspberry Pi. It can talk, nod, and change its ear color, but it has no screen and very little computing power.

The study needed much more than that. Participants have to log in, read instructions, type answers, talk with an LLM, fill out surveys, and sync a Fitbit. And they do all of this at home, on their own Wi-Fi, for six weeks, with no researcher in the room.


πŸ’‘ How DART splits the work

The robot only does what a robot is good at: speaking and moving. The participant's laptop runs a web app that shows the exercise and collects their input. AWS sits in between and handles login, data storage, LLM and voice calls, and messaging to the robot.

When the web app needs Blossom to say something, it sends a short command over MQTT. The robot runs it, reports back when it's done, and the web app moves on to the next step. Because both sides connect out to AWS, it works on any home network without changing router settings.

image
image

πŸ— Stack

  • Next.js 14 and TypeScript, hosted on AWS Amplify Gen 2
  • Cognito for login, DynamoDB and S3 for study data
  • Lambda functions for account setup, the weekly schedule, and email
  • AWS IoT Core (MQTT) between the web app and the robot
  • GPT-4o for conversation, Amazon Polly for Blossom's voice
  • Fitbit OAuth, with tokens stored in AWS Secrets Manager

βš™οΈ What I worked on

I joined in January 2026 and lead the web side of a 16-person research team. I review every PR from the web developers, own the AWS setup, and turn what the PhD researchers need into working features. Most of my work went into the six-week, at-home version of the platform.

Study schedule

Exercises unlock at 4 AM Pacific, missed ones roll over to the next week, researchers can pause a participant, and a weekly Lambda sends a reminder email when someone falls behind. Getting the exercise counter right took a few rounds. In the end I stopped updating the count step by step and computed it from the sessions a participant had actually completed. I also made completing an exercise safe to repeat and wrote a script to fix records that had already drifted.

Researcher portal

Researchers used to open the AWS console to set up a participant. Now they can do it from one page: create accounts, reset passwords, review safety screening responses and flagged chat messages, check survey completion, and start or stop Fitbit data collection.

Chatbot

I reworked how the chatbot remembers earlier sessions, based on feedback from the clinical team. At one point long conversations started failing with 403 errors. Amplify's firewall rejects requests over 8 KB, and we were sending the whole chat history with every message. I changed the server to load the history itself, so each request stays small no matter how long the conversation gets.

Robot sync

I added Replay and Skip buttons for when Blossom crashes in the middle of a session, and fixed a login timing issue that made the web app send commands to the wrong robot channel.

Privacy and data safety

I tightened how sessions are handled on login-only pages, tied every data write to the signed-in account, kept flagged messages from being lost when a session restarts, and removed participant emails from researcher views.


πŸ“Š Results from the paper

  • In-lab study with 103 participants: stress, anxiety, and negative feelings dropped within a single session, and the system got a mean System Usability Scale score of 78.89
  • Six-week home pilot with 4 participants: mean usability score of 87.5
  • The most common requests were talking to Blossom by voice and tighter timing between the website and the robot

The in-lab study ran on the lab's in-lab version of the system. My work was mostly on the six-week home version.


πŸ‘₯ Team

Built with PhD researchers at the Interaction Lab and web developers Jiancheng Li and Qianwei Guo, advised by Prof. Maja J. Matarić.


πŸ”— Links


🧭 Reflection

Most things I built before were used in a demo or for a few weeks. Here, real people used the platform every day for six weeks, and when something broke, a researcher had to reach out to a participant about it. That made me much more careful. Now when I fix a bug, I try to understand why it happened, and I try to build tools so the researchers don't have to come to me for every small change.