Overview
Can you engineer empathy? Root :) is a smart plant pot with an embedded AI agent: it senses the plant’s physiological state (soil moisture, light, temperature), reads the user’s emotional state through voice prosody, and responds with light, facial expressions and gestures that simulate emotion.
My team designed the concept — but instead of assuming which expressive signals would make the pot feel warm and alive, we measured it. Using Kansei Engineering, we ran a 2³ factorial experiment with 62 participants to quantify how three sensory channels (light colour, facial expression, movement dynamics) shape four emotional perceptions.
The headline finding: colour did almost all the emotional work. Movement helped. The facial expression — the channel you’d intuitively design first — barely registered.
/images/root/concept-happy.pngThe Root :) plant pot concept showing a happy facial expression with warm yellow lightThe concept
The project started from the folk saying “talk to your plants so they grow.” We turned that into an interaction model with two directions of empathy:
- The pot advocates for the plant. Sensor data (“thirsty”, “needs light”) is translated into legible emotional displays — colour, posture, expression — that motivate the user to act. Botanical needs are easy to ignore; emotional requests are not.
- The pot listens to the user. Analysing voice prosody (tone, volume, pace), it infers the user’s state and shifts into a listening posture — acting as a confidant in a domestic context often marked by remote-work stress and isolation.
The target users are young adults — students and early-career professionals — who come home to decompress. The pot becomes a “living diary”: you talk through your day while performing small acts of care.
The design question this raises: if the pot’s entire value depends on its emotions being legible, which expressive parameters actually carry that signal? That is a measurable question — so we measured it.
Building the semantic space
Kansei Engineering starts with words, not pixels. We collected ~60 raw emotional terms from two sources: theory (Norman’s emotional design, Picard’s affective computing, basic emotion theory) and two elicitation quizzes where people described a “sad” and a “happy” robot.
We then reduced the list through explicit exclusion criteria — dropping terms that were emotions themselves (sad, happy), abstractions (cry for help), or unobservable concepts (trustworthy, grateful). The 23 surviving terms clustered into three families:
- Affection / Proximity — welcoming, cold, attentive, distant, sociable…
- Energy / Expression — animated, apathetic, calm, agitated…
- Naturalness — human, mechanical.
From these clusters we defined four bipolar 7-point semantic differential scales, each traceable back to its source terms:
| Kansei scale | What it measures |
|---|---|
| Cold — Welcoming | The affective “temperature” of the interaction |
| Dejected — Animated | Perceived energy (arousal) |
| Mechanical — Human | Whether expression reads as alive or artificial |
| Distant — Close | Perceived care and connection |
Why this rigour matters: every scale in the final questionnaire is traceable to either theory or participant language. Nothing was measured just because it sounded nice.
From words to stimuli
On the product side, we mapped the pot’s expressive channels to three manipulable properties, each with two levels calibrated to opposite emotional states:
| Property | Level 1 (“sad”) | Level 2 (“happy”) |
|---|---|---|
| P1 — Light colour | Blue, cool, 40–60% intensity | Yellow, warm, 70–90% intensity |
| P2 — Facial expression | Sad eyes, slow 1000ms fade-in, low contrast | Happy eyes, fast 300ms fade-in, high contrast |
| P3 — Movement | One slow wave (1.5s cycle, small amplitude) | 3–5 fast waves (0.5s cycles, wide amplitude) |
A full 2³ factorial design gave us 8 stimulus combinations (S1–S8), each prototyped as a short video of the pot performing that configuration.
/images/root/concept-sad.pngLevel 1 of P2: the pot with sad eyes, slow fade-in and low contrast/images/root/concept-happy.pngLevel 2 of P2: the pot with happy eyes, fast fade-in and high contrastWhy factorial, not A/B? Testing all 8 combinations lets you isolate the main effect of each channel — you can tell whether warmth comes from the colour, the face, the motion, or an interaction between them. A simple happy-vs-sad comparison would have confounded all three.
The experiment
- N = 62 participants, recruited online. Predominantly young (48.4% aged 18–24, 25.8% aged 25–34), with medium-to-high tech familiarity (M = 3.42 on a 1–5 scale). 56.5% care for plants at home — a near-even split between our target users and outsiders.
- Each participant watched all 8 stimulus videos and rated each on the four Kansei scales.
- Order effects controlled: we built 8 questionnaire versions, each starting the stimulus sequence at a different point (S1→S8, S2→S1, …), rotating the presentation order across participants.
- Informed consent, anonymity, academic context disclosed.
Findings
The prototype produces two clearly distinct emotional states
Averaging the ratings per stimulus revealed two opposite profiles, exactly as the design intended:
| Temperature | Energy | Naturalness | Closeness | |
|---|---|---|---|---|
| S1 — Blue · Sad · Slow (low-arousal) | 2.55 | 2.13 | 3.37 | 3.08 |
| S8 — Yellow · Happy · Fast (high-arousal) | 5.19 | 4.98 | 3.84 | 4.90 |
(Scales 1–7; full means matrix for all 8 stimuli in the report.)
Colour dominates everything
Across Temperature, Energy and Closeness, light colour was by far the strongest factor. Yellow pushed perceptions toward welcoming, animated and close; blue pulled them toward cold, dejected and distant.
The clearest evidence: stimuli combining yellow light with a sad face (S3, S4) still scored high on Energy (4.53 and 4.50) — the colour overpowered the expression entirely.
/images/root/effects-energy.pngChart of factor effects on perceived Energy, showing light colour as by far the strongest factorMovement is a meaningful second channel
Fast, repeated gestures reinforced Temperature and Closeness (though they had a near-zero effect on Energy — surprising, since movement speed is the intuitive “energy” channel).
/images/root/effects-closeness.pngChart of factor effects on perceived Closeness, showing colour dominant with movement as a secondary reinforcementThe facial expression barely mattered
Across all four scales, the sad-vs-happy face produced only residual differences. The channel that feels most “emotional” to design was the weakest signal we measured.
The honest null: nothing made it feel human
The Mechanical—Human scale hovered around the midpoint for every stimulus (3.37–3.84), and all factor effects were close to zero. Colour, expression and movement patterns — at least as we manipulated them — do not make an agent feel alive. The report hypothesises that naturalness lives elsewhere: movement fluidity, micro-pauses, reactivity. That’s a finding, not a failure — it tells the next design iteration exactly where to look.
/images/root/effects-naturalness.pngThe honest null: chart of factor effects on perceived Naturalness, with every factor close to zeroDesign implications
We converted the effect analysis into a sign matrix — a simplified predictive Kansei model that turns results into configuration presets for the agent:
- Empathetic / Welcoming / Close preset: yellow light + fast movement (happy face as optional reinforcement)
- Low-arousal / Calm preset: blue light + slow movement
- Neutral presets: mixed combinations (e.g. yellow + slow) for intermediate states
This is the practical payoff of the method: instead of a designer’s guess, the pot’s emotional vocabulary is now an evidence-backed mapping from internal state to expressive output.
Learnings
- Measure before you polish. We would have spent most of our effort perfecting facial animations — the channel the data showed matters least. One factorial study redirected the design priority to colour and motion.
- Traceability is what makes Kansei defensible. Being able to trace every scale back to participant language or theory made the results interpretable instead of arbitrary.
- Null results are design guidance. The flat Naturalness scale didn’t invalidate the prototype — it scoped the next research question (fluidity, micro-timing, reactivity).
- Order rotation is cheap insurance. Eight questionnaire versions cost us an afternoon and removed a whole class of order-effect objections.