TDK Corporation announces the InvenSense SmartEdgeML^TM, an advanced edge machine learning solution enabling new possibilities for wearables, hearables, AR glasses, IoT, and other products that benefit from machine learning (ML) at the sensor chip level. SmartEdgeML is the first solution to generate and run machine learning models on a 2.5×3mm 6-axis motion sensor IMU at <30µA.

“TDK’s SmartEdgeML is a paradigm shift in edge machine learning, as it will allow developers, ODMs, and OEMs to implement ML-optimized motion sensor algorithms on an IMU sensor chip. This reduces the amount of raw data going to edge processors, which significantly improves device battery life, data privacy, and system latency,” said Sahil Choudhary, Director Motion Sensors and Software at InvenSense, a TDK Group company.

TDK also announces the availability of the InvenSense SmartBug 2.0 (MD-45686-ML), a multi-sensor wireless module consisting of the InvenSense ICM-45686-S IMU. This module works as the perfect evaluation system for users to start with the InvenSense SIF and the ICM-45686-S IMU. The SIF is now available for download, while the MD-45686-ML and ICM-45686-S will be available at distributors by February 1, 2024.

There are three components of SmartEdgeML:

SIF (sensor inference framework): SIF, the software component of SmartEdgeML, is a complete ML framework by TDK. It provides a one-stop-shop for users to collect IMU sensor data, select custom features, build ML models, test ML performance, deploy, and run those models on the ICM-45686-S IMU through the SmartBug 2.0. Tested examples include algorithms such as exercise classification (squats, jumping jacks, lateral raises, or push-ups) and wrist gesture classification (fight, turn, shake, or still).

ICM-45686-S IMU: This is the hardware component of SmartEdgeML. The SmartMotion ICM-45686-S is a 2.5×3mm IMU from the TDK BalancedGyro™ family that enables ML decision tree models to be run on-chip at the lowest current consumption (< 30µA). This new IMU provides premium temperature stability and vibration rejection, making it optimal for applications such as AR glasses, VR, OIS, drones, TWS, and robotics that need a combination of high-performance and ultra-low-power ML algorithms.

SmartBug 2.0 (ML version): MD-45686-ML is an all-in-one multi-sensor wireless module that comes with the ICM-45686-S 6-axis motion sensor and is compatible with the SIF. The small form factor and BLE + USB interface of SmartBug 2.0 allow users to start quickly with SIF so they can move easily from data collection to building ML models, to deploying on the ICM-45686-S IMU. This is the go-to device for getting started with SmartEdgeML.

Main Features and Benefits

Customization: Users can define and customize their own use case and build a motion sensor algorithm with SIF in less than 5 minutes (with AUTO mode). Users can also configure their custom sensor settings, filters and features based on their sensor algorithm requirements.

Ownership: Users can own the data and test the ML model with their own dataset rather than depending on the sensor vendor for data collection.

Time to market: SIF AUTO mode allows an ML beginner to build an ML model in 5 minutes. After the user collects data, the SIF takes care of the rest. Once the model/algorithm is ready and meets the performance criteria, TDK provides an integration guide to run the final algorithm on ICM-45686-S IMU on the user’s system. This end-to-end ML solution on an IMU saves multiple months of algorithm effort.

Ultra-Low Power: The SmartEdgeML solutions can run as low as <30µA. This low power allows the edge processor device to sleep longer and process only smart data coming from the sensors, reducing battery drainage and MIPS cycles.

TDK demonstrates the SmartEdgeML at this week’s Consumer Electronics Show (CES) in Las Vegas. At its booth the demo “SmartEdgeML Machine Learning on a Chip” allows visitors to build a machine learning algorithm on a 2.5×3mm motion sensor in less than 5 minutes.