Embedded AI: Integration of AI algorithms in systems

The challenge: Our client wants to enhance their current system, while accelerating development, by implementing a resource-efficient, modern AI algorithm. The use of AI in the two-wheel vehicle industry presents numerous safety-critical requirements and very limited computing resources on control units, compounded by frequent system changes.

Our solution: To develop non-linear driving functionality, we utilized a closed-loop simulation to train an AI-based control algorithm using reinforcement learning. Setting up an edge device connected to the control unit facilitated rapid prototype testing including real-world retraining to minimize simulation deviations. Deployment and hardware-specific optimization of the trained AI were achieved using the ETAS Embedded AI Coder to ensure compliance with safety standards such as ISO 26262 and MISRA.

Added value for the customer: This easily adaptable simulation and real-world setup, including the ETAS Embedded AI Coder, enabled our client to train, optimize and deploy an AI algorithm while reducing software complexity.

The challenge: The client needs to create a balanced dataset from predominantly homogeneous data for AI training and evaluation. Given the AI’s broad application across various product configurations and environmental conditions, good model generalization is crucial. The model must run on an embedded system with strict hardware constraints, and a comprehensive picture of performance and reliability under real-world conditions is necessary.

Our solution: A customized dashboard displays statistics of the existing data and uses an entropy algorithm to create a balanced dataset for training. Targeted analyses and experiments distil data features relevant for generalization. Based on this, new, adapted data campaigns are planned and implemented. The AI model is specifically optimized for the target hardware and deployed. To minimize testing efforts, shadow mode is employed, where the model is integrated into the end customer’s product and fed real sensor data without influencing AI outputs. Potential anomalies are then uploaded to the cloud, enabling automated AI improvements.

Added value for the customer: An AI that runs robustly for all use cases, leading to cost savings and reliable operation. Collecting relevant data saves unnecessary, expensive data campaigns and generates real added value. Extensive real-world testing fosters confidence in the AI application.