D.O.R.M.A

Inspiration

Mental health concerns among university students have risen sharply, with disclosures increasing from under 1% in 2010/11 to 5.8% in 2022/23. Many students struggle silently, especially late at night when support is limited. We wanted to explore whether an always-available, privacy-respecting robotic companion could provide immediate emotional support and early intervention.

What it does

D.O.R.M.A is a mobile robot assistant designed to patrol university dorms and initiate supportive conversations with students who appear distressed.

Using computer vision and conversational AI, the robot can:

Detect and approach nearby students
Hold natural voice conversations
Recall past interactions to personalise responses
Provide calm, supportive dialogue in moments of stress

The goal is not to replace human support, but to act as an accessible first line of comfort when no one else is immediately available.

How we built it

Hardware

Arduino Uno R4 (motor control and mobility)
Differential drive (two independently powered wheels)
Camera for face detection
Microphone for speech input
Speaker for voice output

Software stack

Python orchestration layer
Facial recognition pipeline for user identification
MongoDB for persistent user memory and conversation history
Retrieval agent to build personalised context
OpenAI LLM for response generation
ElevenLabs for natural voice synthesis

System flow

camera → face recognition → user_id
microphone → speech-to-text
MongoDB → memory retrieval
LLM → personalised response
ElevenLabs → spoken reply

This architecture allows D.O.R.M.A to maintain stateful, evolving interactions rather than stateless chatbot responses.

Challenges we ran into

Reliable identity tracking: Ensuring the system does not treat the same student as a new user required careful handling of facial embeddings and database lookups.
Real-time orchestration: Coordinating vision, audio, LLM calls, and speech synthesis within acceptable latency was non-trivial.
API quota constraints: Managing ElevenLabs credit usage required optimising prompt length and response size.
Hardware–software integration: Synchronising Arduino motor control with the Python AI pipeline required iterative testing.
Data structure design: Storing time-series biometric data efficiently in MongoDB while keeping queries fast needed thoughtful schema design.

Accomplishments that we're proud of

Built a fully stateful conversational robot, not just a stateless chatbot
Implemented persistent memory using MongoDB
Achieved end-to-end voice interaction pipeline
Successfully integrated face recognition with personalised recall
Designed a modular architecture that can scale beyond the prototype

Most importantly, we demonstrated that emotionally aware robotic support in student living spaces is technically feasible.

What we learned

Memory is critical — without persistent context, conversational agents feel shallow
Latency budgets matter more than expected in embodied AI systems
Hardware reliability becomes the bottleneck faster than model quality
Clean modular architecture dramatically speeds up iteration
Even simple personalisation significantly improves perceived empathy

What's next for D.O.R.M.A

Short term

Improve distress detection using multimodal signals (posture, voice tone, physiology)
Add real-time speech-to-text loop for fully continuous conversation
Enhance safety and privacy controls

Medium term

Deploy pilot testing in controlled university environments
Add wellbeing check-ins and trend monitoring
Improve on-device inference to reduce cloud latency

Long term vision

D.O.R.M.A could become a supportive ambient presence in student housing, hospitals, and assisted living environments, providing immediate, personalised comfort while escalating to human support when needed.