Service Overview
Edge AI
FPT Automotive's edge-ai service delivers low-latency, privacy-first intelligence directly on automotive platforms. It integrates driver emotion AI on Qualcomm SA8255P (fusing 20+ CAN signals with sub-100ms inference), hybrid voice assistants on Telechips Dolphin-5 (ASR on NPU, embeddings, and TTS on CPU), and edge computer vision for in-line factory inspection (screw verification in 2s at 99.7% accuracy). Abstracts hardware via Arm NN and ONNX Runtime.
Sub-100ms
Qualcomm emotion inference
1,241ms
Telechips edge voice loop
2 seconds
Edge vision panel inspection
18.7×
INT8 optimization speedup
99.7%
Manufacturing inspection accuracy
Capabilities
Key capabilities
Sub-100ms Inference
Continuous real-time inference meeting strict automotive latency requirements on Qualcomm SA8255P SoC.
INT8 Quantization
Model optimization via INT8 quantization delivering 18.7× speed improvement and 116% throughput gain.
Driver Emotion AI
Fuses 20+ CAN signals to understand driver emotional state and context for in-cabin personalization.
Cloud-Independent Operation
All inference runs on-device — no cloud dependency, no latency penalty, no data transmission.
Automotive SoC Expertise
Deep integration with Qualcomm automotive platforms including native hardware acceleration.
Technology
Technology stack
| Component | Technology | Purpose |
|---|---|---|
| Hardware | Qualcomm Snapdragon, NVIDIA Orin | Edge computing platform |
| Framework | ROS2, TensorFlow Lite | AI model deployment |
| Sensors | LiDAR, Camera, Radar | Environmental perception |
| Communication | CAN-bus, Ethernet | Vehicle network integration |
| Development | C++, Python | Implementation language |
Use cases
Real-world applications
Documented outcomes from actual deployments.
Optimized AI Model for Edge Devices
Resolved QNN conversion pipeline degradation when converting AI models from PyTorch → ONNX → QNN for the Qualcomm SA8255P HTP backend. Root cause was a data layout mismatch during QNN conversion.
Before
QNN conversion produced degraded inference accuracy due to transpose processing mismatch
After
Functional equivalence fully restored on SA8255P HTP backend after graph modification
Edge AI Door Inspection — 2s at 99.7% Accuracy
Edge computer vision system for real-time screw inspection on car door panels at manufacturing workstations. Inference runs entirely on-device, returning immediate pass/fail decisions without cloud round-trips.
Before
3-minute manual inspection per door; growing production volumes made the bottleneck unsustainable
After
2-second edge AI inspection at 99.7% accuracy keeping the production line moving
How we work
Implementation approach
Phase 1: Assessment & Planning
- Analyze current vehicle architecture and sensor configuration
- Identify AI use cases and performance requirements
- Define latency, accuracy, and reliability targets
Phase 2: Model Development & Training
- Develop custom AI models for specific use cases
- Train on automotive-specific datasets
- Optimize models for edge hardware constraints
Phase 3: Integration & Testing
- Integrate models with vehicle systems
- Conduct Hardware-in-the-Loop (HIL) testing
- Validate performance in real-world conditions
Phase 4: Deployment & Monitoring
- Deploy to vehicle fleet
- Monitor performance metrics and system health
- Iterate and improve based on real-world data
Explore more
Related services
Physical AI
Autonomous driving VLMs, robotic manipulation benchmarking, and MLOps on Vertex AI.
Data Analysis & ML
Predictive safety, driver context on Qualcomm, EV diagnostics, and drawing reviews powered by ML.
System Test AI
AI-assisted end-to-end vehicle validation and HIL/SIL simulation compliance.