Disney's Anzellan Animatronics Use AWS Cloud AI to React to Guests

Walt Disney Imagineering R&D has unveiled three Audio-Animatronic prototypes that use Amazon Web Services cloud infrastructure and AI to respond to guests in real time. The figures, named Vilo Vele, Kado Bre'a and Jibu Bre'a, are modeled on the Anzellans, the alien species introduced in The Mandalorian and Grogu. According to The Verge, the animatronics will make their first public appearance at the AWS re:Invent conference in Las Vegas, where attendees can interact with them.
Each figure contains hidden cameras and microphones that capture guest movement and speech. That sensor data is not handled by the robot alone. Multiple AI models run on AWS cloud servers, process the input, and send back a set of behaviors and speech for the figure to perform. The models split the work by time horizon. A fast-decision model handles immediate responses, such as making eye contact to appear attentive, while larger models plan longer interactions across several exchanges.
The most important constraint is that the content is written in advance. Imagineers authored every movement and line of dialogue the models can trigger. The models select from that library. They do not generate free-form speech or improvised motion. Each of the three figures also has its own personality quirks, which gives the models a distinct behavioral profile to work within.
The broader context here is a familiar engineering trade-off. Reflexive behavior and deliberative behavior have different latency budgets, meaning the amount of time a system can take before a response feels wrong. Routing each type to a different model is a practical way to meet both. Keeping the expressive content hand-authored, rather than generated, holds the character voice steady while the selection logic adapts to the guest in front of it.
In my view, the hardest problem is not the language model. It is round-trip latency, the time between a guest's movement and the figure's response. A reaction that arrives a second late reads as a malfunction in a physical character far more quickly than a chatbot's pause reads as thoughtfulness. How the system holds up when several guests move and talk at once will say more about its readiness than any single well-timed glance.
In my view, the privacy dimension deserves the same attention as the animation. Cameras and microphones inside a character that invites conversation change what a guest agrees to when stepping up to it, whether the footage is processed on premises or in the cloud. Any park deployment will need to be explicit about that trade.
In my view, the direction is worth noting. A physical character that can hold a short, contextually appropriate exchange, with behavior shaped by a cloud-hosted model stack rather than a fixed loop, opens interaction designs that scripted animatronics could not support. The re:Invent appearance will be the first chance for attendees to test how well that works in person, and the gap between a demonstration and a busy theme park floor is where the real engineering will be judged.


