1
Customer-
Centric
Innovation
in Automotive
2
Porsche Engineering Company Overview
Engineering Ecosystem & Development Loop
Summary
Use Cases
• Signal Foundation Model
• AI-Corrected Race GPS
• Decoding the Fleet (Agentic Diagnosis)
Agenda
3
Porsche Engineering US is integrated into an international network
of competence centers focusing on digital technologies.
Germany Czech Republic Romania Italy China
• Headquarters of the
International Group at the
Porsche R&D Center Weissach
• Vehicle technologies and
integration
• Function and software
development incl. software
architectures
• Big Data and Artificial
Intelligence
• Testing and simulation
Weissach, Bietigheim-
Bissingen, Leipzig, Mönsheim,
Wolfsburg Prag, Ostrava
• System development and
integration
• Testing and simulation
• Function and software
development
• Big Data and Artificial
Intelligence
• Function and software
development
• Software integration and
software quality
• Big Data and Artificial
Intelligence
• Testing and simulation
• Vehicle testing and
development
• Driver assistance systems
and highly automated driving
functions
• V2X testing
• E-mobility
• Endurance testing and quality
• Driving dynamics
• Connection of virtual and real
testing
• China-specific solutions
• Function and software
development
• Connectivity
• Quality
• Testing and simulation
Cluj, Timișoara Nardò Technical Center Shanghai, Peking
USA
• Market-specific vehicle
functions
• Digital services
• Connectivity and
infotainment
Los Angeles
4
PEUS as Part of Porsche Engineering Group: Company Overview
13 Locations
in Germany, Italy, Czech
Republic, Romania,
China and USA.
LOCATIONS
Founded in 1931 by
Ferdinand Porsche in
Stuttgart
100% subsidiary of
Dr. Ing. . F. Porsche AG
HERITAGE
Employees
PEG Management
EMPLOYEES
Dirk Lappe
(CTO)
Dirk Philipp
(CFO/COO)
Markus-
Christian Eberl
(CEO)
Digital
technologies
Vehicle
expertise
Technologies for the
intelligent and connected
vehicles of the future
FOCUS
5
Our Experience in the Field of Vehicle and System Development
Knows Many Examples.
Complete vehicle development Cayenne Coupé
High-voltage system development and e-mobility
Chassis system development End-2-end electronics architecture
AI in engineering Function development and testing
6
Porsche Engineering Group (PEG) Projects
Porsche Engineering x Sebastian Steudtner
Commercial vehicle development X-BOW with dual-clutch transmission
High-voltage system development 919 Hybrid Chassis Demonstrator
Battery system development Seabob Design and styling of Material Handling
7
No Matter What We Develop, We Always Keep an Eye on the
Integration Throughout The Complete Vehicle. Quality Included.
Highly automated
driving
E-mobility and high-
voltage systems
Artificial IntelligenceConnectivity
ASPICE | ISO26262 | ISO9001 | TISAX
E/E architecture
Drive systems and
functions
Concept
Complete vehicle integration
Chassis systems and
functions
Vehicle body systems
and functions
Development Simulation and testing Series
8
We smartly use our ecosystem to reduce the ‘Time To Car’
Transforming from lab to real
world evaluation in the vehicle
through our Cloud Connected
Realtime Ecosystem
P R O B L E M D E S C R I P T I O N
Controller-centric workflows
→ Long iteration cycles
Offline recordings
→ Hand-operated data offload (USB/loggers)
Manual measurement setup & configuration
→ High effort, low repeatability
Fleet testing at scale
→ Hard to manage, hard to compare results
Research ideas stayed long in lab environments
→ No final Confidence because of missing field tests
9
01 02 03
Cloud Processing / Data Center
Overview of Our Development Loop
EXECUTE & COLLECT EVALUATE OPTIMIZE
AI-enabled Edge Device
Standard Vehicle Architecture
Upload Deploy
DATA L AKE NEURAL NET ANALYZE
10
Slide 11
Technical Components in the Development Loop
On-premise ResourcesCloud ResourcesEdge Devices
TRAINED
MODELS
& DATA MODELS
Sensor
Data &
Model
Results
NVIDIA DGX Systems
with Sharding Capabilities
Workload Management
Centralized Data Lake
Experiment Tracking & Model Registry
Scalable Processing Platform
IoT Edge Platform with Fleet Monitoring
Test Vehicle equipped with additional
Edge Device Capabilities
Cloud-connected Vehicle Middleware
Azure
Azure IoT Hub
12
Start at the Vehicle: the Edge Device as the Data Source
Devices
13
The CarDataBox Serves as Edge Device in Our AI Ecosystem.
Software
Application Layer
Connected Vehicle Middleware
Custom Hardware Platform
powered by NVIDIA Jetson
Hardware
14
Technical Structure of the CarDataBox
Core Software
Layer
Hardware Layer
Application
Layer
Connected Vehicle Middleware
Custom Hardware Platform (NVIDIA Jetson)
Core
Core Runtime
CAN Interpreter Automotive Ethernet FlexRay Interpreter
Camera Interpreter Application ManagerConfiguration Management
ROS interface
Machine Learning
Component
ROS interface
Real-Time
Application
ROS interface
Azure Uploader
Component
. . . . .
Docker
15
Remote Application Deployment on Edge Devices
Engineer Container
Registry
Remote
Deployment
Data
Lake
Analytics-, ML- and
Processing-Pipelines
Weight
Updates
DEVELOP & DEPLOY EXECUTE & COLLECT ANALYZE & OPTIMIZE
EDGE DEVICE
ADAS App ML App Evaluation App
Azure
16
Technical Components in the Development Loop
On-premise ResourcesCloud ResourcesEdge Devices
TRAINED
MODELS
& DATA MODELS
Sensor
Data &
Model
Results
NVIDIA DGX Systems
with Sharding Capabilities
Workload Management
Centralized Data Lake
Experiment Tracking & Model Registry
Scalable Processing Platform
IoT Edge Platform with Fleet Monitoring
Test Vehicle equipped with additional
Edge Device Capabilities
Cloud-connected Vehicle Middleware
Slide 17
18
Technical Components
in the Development Loop
Resources
DockerAzure
Azure IoT Hub
19
Processing the Collected Data in the Cloud
NVIDIA B200
on-premise
Large Model Training
Kubernetes
Spaces
Data Lake
Indexed by
• Vehicle signals
• Vehicle derivate
• Locations
• Weather
• …
Project
space 1
Project
space 2
Project
space 3
Curated
Datasets
Fleet
Analytics
Small Model Training
and Model Inference
AzureAzure IoT
Hub
20
Technical Components in the Development Loop
On-premise ResourcesCloud ResourcesEdge Devices
TRAINED
MODELS
& DATA MODELS
Sensor
Data &
Model
Results
NVIDIA DGX Systems
with Sharding Capabilities
Workload Management
Centralized Data Lake
Experiment Tracking & Model Registry
Scalable Processing Platform
IoT Edge Platform with Fleet Monitoring
Test Vehicle equipped with additional
Edge Device Capabilities
Cloud-connected Vehicle Middleware
Slide 21
Slide 17
22
Technical components
in the development loop
Resources
23
NVIDIA DGX Platform
Local Inference
Workload Scheduling & Model Training
Containerization
Partitioning of GPUs to suit workloads
DGX Cluster (B200)
MiG Partitioning
Docker / NVIDIA Container Runtime
SLURM + PyTorch
Deep Learning
Engineers
Agent
Developers
Tools /
Products
PoC ServicesLLM Inference Model Training
Model deployment,
experiment tracking
LLM Evaluation/
Monitoring
NIMvLLM / Ollama
24
Efficiency Gains From our Cloud & AI Ecosystem
An end-to-end, consistent ecosystem spanning data collection,
real-time validation, scalable training, and edge deployment
Vehicle architecture-independent software development enabling
flexible adaptation across platforms
Fleet-scale operation, evaluation, and data collection
Direct in-vehicle evaluation & Updated within minutes
Unified global AI model lifecycle, experiment tracking, & governance
DockerAzure
Azure IoT Hub
25
Applying our Ecosystem to Real-World Challenges
USE CASE #1
AI-based scenario, pattern
and anomaly finder
Improving vehicle positioning
accuracy during high dynamic
driving
AI-Agent driven vehicle data
analysis
Signal Foundation Model Race GPS Vehicle Data Analysis
USE CASE #2 USE CASE #3
26
USE CASE #1
Signal Foundation Model
AI-based scenario, pattern and anomaly finder
27
Signal Foundation Model
for ADAS Scenario Search, Clustering, and Description
C H A L L E N G E
• Large amounts of unlabeled signal recordings
• Rare/safety-critical events are hard to find, compare,
and quantify across fleets
A P P R O A C H
Clustering existing database of signal recordings into
searchable database of scenarios
Latent space
I N P U T
Driving scenarios
as time series data
O U T P U T
Searchable latent
space
28
Applying Our Ecosystem
DEVELOP & DEPLOY ANALYZE & OPTIMIZEEXECUTE & COLLECT
Signals Recorded
• Lane Detections
• Object Detection
• Acceleration/Speed
• …
Analytics-, ML- and
Processing-Pipelines
Data
Lake
Weight
Updates
Engineer
Azure
29
Signal Foundation Model
* Embedding – the coordinates of a single point in the latent space | ** Description – the label of the current scenario represented by the embedding
• Extract the important features of the scenario
• Represent it by a set of vectors called “embeddings”
• Embedding is a collection of
coordinates in a multidimensional
space
• Similar scenarios are residing
in the same neighborhood
in this “Latent space”
Description
• Similarity searching using
embeddings
• Generate similar scenarios
using embeddings
• Similarity search using text
• Forecasting classification
Use Cases
Latent space
Time Series “Foundation Model” Large Language ModelTime Series Data
Domain
input signals Latent Space
Encoder Decoder
Distribution
µ
σ2 Sample
Embedding*
Description** [label] A car at a medium
distance in the middle lane …
Softmax
Linear
RMS Norm
Feed Forward SwiGLU
RMS Norm
RMS Norm
Embeddings
Self-Attention
(Grouped Multi-Query Attention) with KV Cache
α κ ν
30
Signal Foundation Model
Initial application to ADAS
given a provided scene,
search for similar ones in a database
given an encoded scene,
generate signals for similar situations
given an encoded scenario,
describe in natural language
Search Scenes Generate ScenesDescribe Scene
31
Signal Foundation Model
Benefits
Turn raw recordings
into a searchable scenario database
Faster root-cause analysis
through similarity search
Generate scenario variants
to expand test coverage
Closed-loop improvement from
deployment to optimization
32
USE CASE #2
AI-Corrected Race GPS
for high-dynamic driving
33
High Dynamic Driving Reduces GPS Accuracy
C H A L L E N G E
• Road car GPS noise: 5–10 m uncertainty
• Navigation systems leverage map matching & plausibility checks
• Classical GPS corrections fail in high dynamic driving situations
G O A L
• Predict and correct GPS position error directly from vehicle sensor
data using machine learning
• Achieve accuracy close to high-precision DGPS
Track from Standard GPS
Track from High Precision GPS
34
Solution
LSTM to increase GPS accuracy based on series signals
DEVELOP & DEPLOY ANALYZE & OPTIMIZEEXECUTE & COLLECT
Data Lake
Engineer
Series Signals Recorded
• High-precision GPS
• Steering / Yaw Angle
• Acceleration / Deceleration
• IMU (Inertial Measurement Unit) EDGE DEVICE
Weight Updates
Analytics-, ML- and
Processing-Pipelines
Azure
35
GPS Accuracy Model
AI-Corrected GPS
• AI based Sensor Fusion to improve
accuracy based on dynamic features
• Considers time history of vehicle
trajectory
• Targets low resource consumption to
allow in vehicle processing
Description
LSTM
Model
INPUT OUTPUT
High Precision GPS
Standard
GPS
GPS-specific loss function
36
Benefits: Increased GPS Accuracy Through LSTM
Accuracy boost:
Achieves a ~95% error reduction
GPS 5–10 m → ~1–2 m (~ m RMSE)
Real-world proof:
Validated on different cars (., Porsche 911)
→ track-quality racing line & lap timing
Based on standard vehicle bus signals
high-precision GPS ground truth
(training on DGX, edge-ready)
Track from Standard GPS
Track from High Precision GPS
AI-corrected GPS Position
37
USE CASE #3
Decoding the Fleet
AI Agent-Driven Vehicle Data Analysis
38
AI-driven Development Fleet Insights Reduce Time to Reaction
C H A L L E N G E
• Large amounts and many modalities of vehicle data
• Ambiguous relationships between diagnosis information
• High efforts around domain specific analyses
A P P R O A C H
• Anomaly detection on a diagnosis protocol level
• Large Protocol Transformer–Decoder on development
fleet protocols
• High level of automation with modular AI workflows
39
Our solutions integrate into the development data loop
and enable early diagnosis.
Development fleet vehicles
Issue resolution / development measures
Data lake Ml-driven exploration
& detection
Timeseries anomaly detection
Data feedback loops
Low-code
workflow app
Customizable
analysis insights
Agentic AI Analysis
40
Multi Agent Expert Process Provides Granular Diagnosis Insights
ADAS agent
Chassis agent
Battery agent
…
Root cause
diagnostic insights /
development measures
Exploration &
detection agent
Diagnosis
protocol
41
Benefits of the Agentic Diagnosis Insights
Managing complexity
Efficient handling of
large data volumes
Clarity
Information and causal relationships
at a glance
Enhance human expert knowledge
Multi-Agents provide insights where user
knowledge is limited
Time savings
Fast analysis of diagnosis information,
first hypotheses in minutes vs hours
Porsche Engineering North America 42Public
43
Faster ‘Time to Car’ With a Cloud-Connected Real-Time AI Ecosystem
from lab workflows → in-vehicle evaluation at fleet scale
Why change What we built Why it matters
Controller-centric + offline workflows
create long iteration cycles and high
manual effort. Fleet testing is hard to
manage and compare.
A repeatable loop connecting vehicle edge
→ cloud data lake → on-prem training →
deployment back to the vehicle, including
tracking/registry and workload
management.
Signal foundation model
turns raw recordings into a searchable scenario database
→ faster root-cause analysis and scalable safety
quantification.
Race GPS
improves positioning from 5–10m to ~1–2m
(~95% error reduction).
Agentic diagnosis
First hypotheses in minutes vs hours, reducing
dependency on scarce expert knowledge.
44
Porsche Engineering
Benedict Kistner
Digital Solutions, Porsche Engineering US
@
Leon Eisemann
Machine Learning Engineer, Porsche Engineering US
@
Daniel Schumacher
Cloud Architect, Porsche Engineering
@
mailto:@
mailto:@
mailto:@
mailto:@
mailto:@
mailto:@
mailto:@
mailto:@
mailto:@
