As I write this, Embedded World 2023 is nearly upon us. SensiML will be in Nuremberg once again showing our latest toolkit features and exciting new demos. This year our demos include a keyword spotting exhibit that exploits speaker dependency to demonstrate a voice ID Smart Lock, a full end-to-end TinyML model development demo showing the ease of model building, and a new ‘Wizard Wand’ smart toy demo and game contest which is the focus of this blog.
Smart toys present a unique set of challenges for TinyML compared to other applications. Most of these are driven by the need for aggressive cost and rapid development as toy developers must deliver compelling features that will delight users while also fitting within tight channel fulfillment timelines and rigid price points. Thus, TinyML models as applied to toys must deliver on a combination of conflicting attributes:
- Low power – To minimize battery cost
- Compact footprint – Enabling devices that can be delivered in the smallest MCU memory SKUs
- High accuracy – To deliver on the user experience
- Rapid train/test cycles – To hit manufacturing deadlines for make-or-break holiday channel fulfillment
The SensiML Wizard Wand game showcases how IoT sensor recognition model(s) can serve as a central feature for developing smart toy and gaming applications. In the Wizard Wand example, we use an integrated IMU sensor to measure motion of a toy wand incorporating a series of trained gestures into a recognition model. These gestures and the means for detecting a user’s ability to replicate them serve as the basis for the game. Players are rewarded each time they match the correct gesture in a prescribed sequence. Completing all such gestures within a given timeframe then serves as the overall game objective with basic UI, LEDs, sounds, and wireless event notifications to companion desktop and mobile applications as the remaining elements.
Watch the video clip below to see an example of the game in action:
So how does the model stack up in terms of resource usage? From the profile data generated in SensiML Analytics Studio, we see that the resulting model occupies less then 30kB of SRAM and only 7kB of flash, making it suitable to ultra-low cost MCUs as commonly found in toy products.
To learn more about this demo, visit the Wizard Wand application example documentation. For the actual firmware (based on the M5Stack M5StickC-PLUS IoT Dev Kit), see the details provided within our M5StickC PLUS platform documentation.
And if you're in Nuremberg for Embedded World 2023, come stop by our booth (Hall 4-238) and give the game a try yourself. You just might win an M5StickC PLUS of your own!