Lumi

An AI robot that monitors anger patterns, interrupts escalation in real time, and builds empathy through light, sound, and voice.

AI Social Robot Anger De escalation Physical Product
Lumi robot, three views showing green default state, blue LED band while listening, and yellow/red LED band during anger detection

At a Glance

Role: Prototype Lead and Lead User Testing Researcher
Timeline: Ten week graduate studio at SCAD, two testing sprints
Platform: Physical AI social robot for the home
Outcome: Two testing sprints with 15 participants validated sound effectiveness and shaped the final voice design

A robot that turns down the heat

Lumi is a social robot that helps users visually monitor their pattern of anger, cool them down during ongoing arguments, and facilitate greater self awareness in their personal relationships.

Using NLP, facial expression recognition, and SLAM technology, Lumi detects arousal levels and emotion changes during arguments, responding with a precise, two stage intervention designed to stop escalation and open the door to reflection.

10 Week Project
15 Users Tested
2 Test Sprints

My Role

Prototype Lead · Content Writer · Product Designer for slide decks · Lead User Testing Researcher

Technologies

NLP (Natural Language Processing) · Facial Expression Recognition · SLAM (Simultaneous Localization and Mapping) · Amazon Alexa Voice SDK

Anger without a way out

Anger is a core human emotion that people constantly struggle to keep at bay. Younger adults, ages 16 to 24 are especially prone to losing control, with less impulse regulation and fewer emotional tools.

Problems with anger lead to overly aggressive behavior that negatively affects lives and the people around them. Users are unable to express feelings without getting emotional, struggle to recognize their own triggers, and lack specific methods to stop escalating anger before it causes damage.

How Might We

Help users manage and deescalate anger in real time, improve their interactions with the people they live with, and support reflection on their emotional patterns using an AI powered robot?

HMW breakdown: support users to improve their relationship with people living together using an AI robot — sub-questions covering stopping escalation, monitoring anger patterns, cooling down in the moment, and designing coping methods

How Might We breakdown, four design sub questions shaping the brief

Target Audience

Ages 16 to 24. Younger adults have a harder time managing anger and impulse control, making them the population that could benefit most from Lumi's presence in their homes.

Proto Personas

Megan
19 · Atlanta, GA · College student, lives with boyfriend

Goal: have conversations without them turning into arguments.

Frustration: arguments escalate before she realizes they have started, leaving her feeling guilty and misunderstood. She wants a way to catch the moment before things go too far.

Max
22 · Los Angeles, CA · Lives with family

Goal: be more understanding and not get angry so easily.

Frustration: he knows his triggers but cannot stop himself in the moment. Family arguments leave him feeling ashamed after the fact, and he has no way to track whether he is actually improving over time.

Proto Persona 1 — Megan, 19, Atlanta GA. Behaviors: long-winded rants, hard time admitting fault. Pain points: frequent fighting strains relationship, can't communicate frustration without getting emotional. Goals: find middle ground during arguments, stop the building anger.
Proto Persona 2 — Max, 22, Los Angeles CA. Behaviors: gets very angry with close people, short fuse, uncontrollable rage. Pain points: harbors resentment toward brother, unsure how to control anger. Goals: better control anger around family, seek outlet for redirecting rage.

Megan and Max, detailed proto persona profiles grounding the target audience

Defining the direction

Extensive secondary research shaped the target range, product type, and AI integration approach, narrowing a broad emotional problem into a feasible, testable design direction.

The ten week plan anchored the process around two user testing sprints.

Secondary research narrowed the concept toward a home based AI social robot, and known limitations were acknowledged before concept development began.

Hypothesis: anger control problems can be minimized in ages 16 to 24 through conflict resolution methods; AI robot prevents further escalation and provides coping methods; we will know it worked when arguments are successfully resolved

Design hypothesis, the foundation for both user testing sprints

Goals

Three design goals emerged from the research phase and guided every subsequent decision. They ran in sequence.

  • Stop the escalation of the anger
  • Build an empathetic relationship with the robot
  • Understand the patterns of anger within themselves and others

Designing the system

Lumi's concept is a two stage system: first interrupt the escalation, then facilitate emotional reflection through Cognitive Reframing.

Proposed Solution: Lumi is a social robot detecting anger level and responding via color-changing LED lights, unexpected sounds, and CBT-based AI voice assistant

Proposed solution, a two stage physical robot using sound, light, and voice

Lumi positions itself as a neutral mediator between both parties.

How Lumi works

Research into existing products and future trends in AI led the team toward three core technologies: NLP for semantic and tonal cues, facial expression recognition for visual emotion signals, and SLAM technology (inspired by Roomba) to navigate the home and approach users quickly.

How Would Lumi Work: NLP for semantic cues, SLAM technology for navigation, and facial expression changes to detect anger
AI Input/Output diagram: Input (arousal level, facial expression, non-verbal cues, semantic verbal, location) processed by AI (NLP, SLAM, facial expression recognition), Output (stop further escalating the anger)

Sensor inputs flow through the AI layer to produce targeted responses

Low Fi Prototype

Low-fi prototype sketch: annotated diagram showing camera on spinning head, speaker, LED light strip, and wheels for mobility in all directions

Low fidelity prototype, core hardware components annotated on an early sketch

Use Cases

Two use cases were designed to address the full range of anger intensities. At low anger, subtlety is paramount, an unobtrusive cue that doesn't break the conversation. At high anger, the intervention must be immediate and impossible to ignore.

Use Cases: low anger, subtle sound to unobtrusively alert. High anger, unexpected response for immediate halt in escalation.
Use Case 1: Low Level of Anger, low anger detected, subtle sound, unobtrusively alert the person without stopping the conversation. Use Case 2: High Level of Anger, high anger detected, unexpected response, immediate halt in escalation.

Use Case 1 (low anger) and Use Case 2 (high anger), matched responses by intensity

Core Features

Six capabilities make the two stage system work.

  • Speaker · provides a source of sound to cool users down in heated moments
  • Microphone · detects the noise during an argument and locates the users in their environment
  • Camera · the head spins 360 degrees and the camera detects facial expressions
  • SLAM · maps the environment of the house
  • LED light · warns the users with different colors of light depending on their level of anger
  • Mobility · quickly approaches the users and provides a cool down method during a heated argument

Physical Design

Lumi's form is futuristic yet friendly, inspired by BB8 and designed to be simple and non threatening. Brand values: Friendly, Warm, Empathy. The LED light band is the primary emotional display, changing color in real time to signal the current anger level.

Green, Default state, no anger detected
Blue, Detecting voice
Yellow, Low anger detected
Red, High anger detected

Lumi's Persona

To ensure consistent character across all interactions, the team built a brand persona for Lumi: approachable, empathetic, and friendly, with a calm and warm tone of voice.

Two Stage Concept Direction

Stage 1: Stop escalating the anger, alert using jingle sound and light at low anger; unexpected sound sources at high anger. Stage 2: Coping Method via Cognitive Reframing, alternative cognitive thinking to restructure perception positively; three response methods: mirroring, restatement, and effective response.

Stage 1 interrupts; Stage 2 reframes

Scripting empathy

After mapping common argument scenarios, three voice scripts were created to cover the emotional states a user might be in when approaching Lumi, and to train Amazon Alexa's intent recognition on realistic conversation data.

Lumi Voice Flow diagram: full branching conversation map showing paths for users with negative feelings wanting to talk, users who are unsure, and users who don't want to talk right now

Full voice flow, branching paths cover all three emotional entry states

Lumi in action

Two storyboards place Lumi in realistic domestic scenarios, grounding the concept in the daily lives of the proto personas and communicating the two stage response model.

Storyboard 2: Response to High Anger, Max and his brother are yelling about a gaming console. Lumi rushes over and plays an unexpected humorous audio clip, a funny voice filter or a nostalgic sound, chosen to break the emotional spiral through surprise rather than confrontation. Sean hands over the console and apologizes.

Storyboard 2, Lumi's high anger intervention: unexpected humor breaks the tension immediately

Two sprints, real insight

Two rounds of user testing across 15 participants validated sound effectiveness and refined the voice design. Each sprint combined quantitative surveys with contextual inquiry sessions.

Evaluation Questions

Sprint 1 set out to answer five concept questions and two design questions.

  1. Does the interruption from the AI robot with an unexpected sound actually cool them down during an argument?
  2. What kind of sound most effectively gets the user's attention?
  3. How do users perceive the unexpected sound effect during the argument?
  4. How do users feel about getting different responses for different stages of anger?
  5. Do they want consistency with content or do they want varied audio from the robot?
  6. Do users feel that the size and design of the robot would be effective in calming them?
  7. Is the movement and speed of the robot enough to catch the user's attention during a heated argument?

MVP Testing Plan

MVP Testing Plan: 6-step process — recall the angry moment, write about the past conflict, Wizard of Oz simulation (understand sound effect and movement), quantitative survey, explore shape and size of the robot, contextual inquiry and evaluate

MVP testing plan, six step Wizard of Oz simulation run across 8 participants

Sprint 1, Sound Effectiveness

8 participants · Quantitative survey (20 min) + Contextual inquiry (30 min)

Results: animal sounds were most effective for low anger situations; playing back the user's own voice was most effective at high anger, both surprising and disarming.

User Testing Results: Sprint 1, 8 participants — five dimensions evaluated: Sound, Speed, Lights, Size, Movement

Sprint 1, five dimensions evaluated across 8 participants

Sound results: Low — animal sounds, nature sounds, short jingle. High — funny voice filter, record of user's voice, nostalgic music
Sound results bar charts: Low Anger — animal sounds rated highest (3.8/5); High Anger — user's own voice rated highest (4.5/5)

Sound results, animal sounds top for low anger; user's own voice most effective at high anger

Movement results: Low — no movement preferred; High — users want active movement like dancing or head spinning
Size results: half preferred smaller 7-inch size (fits small room, cuter); half preferred bigger 11-inch (more presence, reminds them of a pet)

Movement and size, active motion preferred at high anger; size split evenly between 7″ and 11″

Sprint 2, Voice Design Naturalness

7 participants · Contextual inquiry (30 min)

Lumi's voice scripts were tested via a foam physical prototype and the Amazon Alexa developer environment, simulating realistic interaction scenarios.

Building Out Sample Code: leveraged Amazon Alexa developer tools to integrate voice scripts; conducted user tests with a physical foam prototype of Lumi

Voice interactions built in Amazon's developer environment, tested with a foam prototype

Coding: Amazon Alexa Intents editor and JavaScript code for FeelBadIntent and DisruptiveListenerHandler

Amazon Alexa Developer Console, voice intents coded and tested in real time

Key Insights

Key Insights: Formalness, users wanted less formal dialog, felt too much like therapy. Reframing, users wanted Lumi to mirror more of their words. Root, more focus on anger triggers. Empathy, needs kinder, more considerate responses.

Four dimensions of voice design feedback from Sprint 2

Formalness

Users wanted a less formal dialog, it felt too clinical, like therapy. Lumi needed a warmer, more casual tone.

Reframing

Users wanted Lumi to mirror more of their words back to them, creating a more personal and connected feel.

Root Cause

More focus on identifying the underlying triggers of anger, rather than addressing only the surface level response.

Empathy

Responses needed more consideration for what users were actually expressing, kinder, gentler language throughout.

Design Changes

Changes slide: expanded sound options for low and high anger, LED flash behavior refined, and the size progression from the tested 7 and 11 inch options through a 10 inch iteration to a final 8 inches

Post testing changes: expanded sound library, refined LED flash behavior, and a final size of 8 inches, settled after testing the 7 and 11 inch options

Validating a new category of emotional intervention

This semester long project was completed in a graduate UX studio at SCAD. It demonstrated that a physical AI robot using sound, light, and voice could effectively interrupt and deescalate arguments, with strong, positive user response across both testing sprints.

Key Finding

Unexpected audio stimuli outperformed every other interrupt method tested, and users who engaged with the Cognitive Reframing scripts reported feeling genuinely heard rather than managed. The long term pattern recognition model was the feature users most wanted to see built, validating the core premise that emotional support tools need to improve the more they know you.

Key Takeaways

Two sprints of testing across 15 users confirmed that unexpected audio stimuli and Cognitive Behavioral Therapy methods are not just accepted by younger adults but actively preferred over conventional calm down prompts. The pattern recognition model proved compelling to users who had never considered that a device could learn their specific triggers over time. The biggest design lesson was that emotional tools need to feel like a relationship, not a feature. Users responded to Lumi not as a product interrupting an argument but as a presence that understood them.

Future Directions

Onboarding

A personalized onboarding experience to build a custom relationship between Lumi and each individual user from day one.

Companion App

A mobile app that lets users review anger patterns over time, track emotional growth, and manage Lumi's settings.

OutcomeTwo testing sprints with 15 participants validated sound effectiveness and shaped the final voice design.

Next project WonderBudi An AI robot that calms children during emotional distress, helps teachers recognize a child in need, and connects families and counselors to support.

Open to full time roles and select freelance work. Say hello, find me on LinkedIn, or use the contact page.