Amazon's Astro Designer Reveals Why Robot Character Must Come First
A moving robot is already a character — whether you designed one or not. Amazon's lead UX Sound Designer for Astro spent years learning that lesson the hard way, and the playbook he built applies to every embodied AI product shipping right now.
The story
In 2018, Amazon hired Mike Forst to define the sound and voice of Astro, its first consumer home robot. What he found was a team split on a foundational question: was Astro just Alexa on wheels, or its own character? Forst and most of the UX team pushed for the latter — and user testing proved them right. People didn't perceive Astro as Alexa; they saw a distinct personality, and Alexa's voice on the device felt "strange and creepy." The fix: Alexa handled speech as a supporting role, while Astro communicated through sound, motion, and facial expressions as the lead.
The core insight is deceptively simple: character is a design system, not a personality sprinkled on at the end. Every decision about how Astro moved, paused, or reacted was a character decision. Without a defined character foundation, those decisions get made by default — inconsistently, by whoever happens to be in the room. Users feel every seam, even if they can't name it. Forst points to Astro's "Sing" sequence as a concrete failure: the expressive moment works, but without transitional "character stitching" before and after, it reads as a clip playing on a robot rather than something coming from within.
The wake-up sequence is the counter-example. Written first as a story — Astro waking in a new home, eager to join a family — it then got a sound vocabulary, and only then did animators build motion and expressions around that emotional arc. Users in testing described it as "alive." That word is the target.
The honest admission: Amazon didn't fully get this right. On a large team racing deadlines, character is always quieter than the things that break. It's easy to defer. The cost only becomes visible in products that feel technically impressive but lifeless — or worse, untrustworthy.
With embodied AI accelerating fast, Forst's ask to product leaders is blunt: define character before you define interactions, build story and sound into the character pipeline (not the production pipeline), and design for adaptation — because the products that will matter are the ones that deepen through use, not just ones that stay consistent.
Reality meter
Why this score?
Trust Layer Defining a robot's character as a foundational design system — before interactions, motion, or sound are built — is the critical missing practice in embodied AI product development.
Defining a robot's character as a foundational design system — before interactions, motion, or sound are built — is the critical missing practice in embodied AI product development.
- User testing showed people did not perceive Astro as Alexa; Alexa's voice on the mobile robot felt 'strange and creepy,' validating the decision to give Astro a distinct, non-verbal character identity.
- Astro's wake-up sequence — written first as a narrative, then scored with sound, then animated — was described by early testers as 'alive,' with sound, motion, and facial expressions expressing the same character in harmony.
- The 'Sing' sequence is cited as a concrete failure: the expressive moment works in isolation, but missing transitional 'character stitching' makes it read as a clip playing on the robot rather than an expression of its character.
- Astro's emotional range was deliberately constrained by design — engineered to avoid extremes of sadness or anger and to resolve reactions on a positive note — as an explicit trust-management decision.
- Forst states that contextual character adaptation worked well when situations were pre-designed, but a real home generates more situations than any team can predict, leaving a long tail of moments the system was never prepared for.
- The account is a single practitioner's first-person retrospective with no independent corroboration from other Astro team members or published user-testing data.
- Claims about what 'industry is missing' are asserted broadly without citing specific competing products or comparative research — the generalization may not hold across all embodied AI development contexts.
The core claims are grounded in a named, shipped product with referenced user-testing outcomes — not speculation — though the retrospective framing means failures are self-reported and unverified.
The piece is notably self-aware about what Astro got wrong and avoids overclaiming; the forward-looking AI statements are framed as potential, not certainty, keeping hype low.
The framework directly addresses a structural gap in how embodied AI products are built today, making it immediately actionable for any team shipping a robot or conversational agent — the timing with the current wave of embodied AI is high-relevance.
- 1 source on file
- Avg trust 40/100
- Trust 40/100
Time horizon
Community read
Glossary
- uncanny valley
- The unsettling feeling people experience when robots or digital characters look or behave almost—but not quite—like humans, creating discomfort rather than familiarity.
- embodied AI
- Artificial intelligence systems that exist in physical form (like robots) and interact with the physical world, rather than existing only as software or voice interfaces.
- character stitching
- The seamless transitions between different expressive states or moments in a character's behavior; poor stitching makes actions appear disconnected rather than part of a continuous performance.
- state machine
- A system that moves between distinct, predefined states based on specific conditions or inputs, often appearing mechanical or rule-based rather than natural or adaptive.
- RL-based (reinforcement learning-based)
- Systems that learn and improve behavior by receiving rewards or penalties for actions, allowing them to adapt and optimize responses without explicit hand-authored rules.
- character coherence
- The consistency and believability of a character's personality, emotional responses, and behavior across different situations and interactions.
What's your read?
Your read shapes future topic weighting.
Your vote feeds topic weights, community direction and future prioritisation. Open community direction
Sources
Optional Submit a prediction Optional: add your prediction on the core question if you like.
Prediction
Will the majority of new consumer embodied AI products launching in the next two years define a formal character system before shipping, rather than treating personality as a post-launch iteration?