Boosting confidence in tiny AI deployment in embedded control systems
summary
Claims a path to replace physical sensors with virtual sensor models (hands-on-wheel from steering-gear telemetry) and to swap decades of ultrasonic "spaghetti code" for a single neural model that can be reparameterized across vehicle variants. A Bosch-derived tool called embedded AI coder generates numerically equivalent C for constrained MCUs, emits test suites and safety docs, and reports concrete benefits: per-layer cycle counts, ROM/RAM sizing, ~33% fewer memory loads and up to 50% RAM reduction, deployed on platforms with billions of OS instances.
AI-written summary · about
chapters (11)
- 0:00EAS and AI in embedded systems
- 1:36AI benefits for real-time embedded
- 3:41Replacing sensors with virtual ones
- 5:26Data scientists vs embedded engineers
- 6:34Safety and trust in AI-generated code
- 9:05Correct code generation proof
- 11:13Reparameterization for multi-product
- 12:40Model choice for embedded constraints
- 15:08Neural network optimization
- 17:19Memory safety and resource efficiency
- 18:24Industrial applications and tools
related talks
08mHuman–AI Collab for Reliable Automotive Control Systems
52mPhysical AI: Crafting Resilient Systems with Emotional Intelligence
42mThe Stakes of Green AI for Embedded Systems
20mDriving Intelligence: The Software Shift Shaping Tomorrow’s Cars
04mMultiple embedded and Cloud AI solutions at ST
18mTransforming MCUs in an AI-Defined Era
57mWriting Safety Critical Automotive Software for High Perf AI Hardware
22mAI On The Farthest Edge