Interview with Steve Stoddard, Sonatus
“AI model inputs can be tagged to VSS signal names upfront”
Steve Stoddard brings experience in scaling AI across distributed vehicle architectures. At ACC US 2026, he will outline how consolidated compute platforms, abstraction layers and robust toolchains can accelerate the deployment of intelligent, adaptive vehicles.
Sonatus
Scaling edge AI across vehicle platforms remains a complex challenge. From compute constraints and legacy architectures to abstraction and toolchains, Steve Stoddard, Principal Product Manager AI at Sonatus, explains where scalability breaks down and what OEMs must prioritise next.
As vehicles evolve into software
defined and increasingly AI driven systems, deploying intelligence at
the edge is becoming a decisive architectural challenge. Automakers must
balance model complexity, compute headroom, safety requirements and cost
efficiency, all while working across heterogeneous platforms and legacy
constraints.
Steve Stoddard, Principal Product Manager, AI at Sonatus,
focuses precisely on this intersection of architecture and scalability. In his
upcoming Automotive Computing Conference US 2026
presentation, Beyond Autonomy: Scaling Edge AI for Smarter, Safer Vehicles,
he explores how edge AI can be deployed efficiently across general purpose
ECUs, gateways and domain controllers. Ahead of the conference, we had the
opportunity to speak with him.
ADT: What architectural challenges arise when
scaling edge AI across vehicle platforms?
Stoddard: The biggest challenges arise from platform
heterogeneity. Each platform has its own mix of ECUs, SoCs, sensors, data
signals, and message definitions. An AI model that runs cleanly on one vehicle
may be non deployable on another because the target ECU lacks compute headroom,
cannot access the required data inputs, or uses different message naming. Even
when the logic is portable, AI models often require calibration for different
vehicles and may require engineering work to re configure for different message
naming, signal parameters, or timing. This fragmentation turns every deployment
into a bespoke integration effort rather than a scalable capability.
How do OEMs balance intelligence, compute constraints and
cost in real world deployments?
In vehicle AI workloads are often segregated by domain. ADAS and other safety critical controls are allocated
to dedicated ECUs during architecture design and core function definition.
Other AI features, such as comfort, personalisation, and convenience features,
are often relegated to shared resources on infotainment ECUs or other high
performance computing platforms when headroom allows. By the mid to late stages
of vehicle programmes, that headroom is usually exhausted. To work around
compute constraints, it is common to define intelligent features using simpler,
calibrated algorithms with lookup tables for the applicable operating domains.
What role does software abstraction play in accelerating
edge AI adoption?
The abstraction of CAN or Ethernet networks and signal
definitions enables greater portability of developed AI models from one vehicle
platform to another. For example, AI model inputs can
be tagged to VSS signal names upfront. VSS signal to CAN mappings can then
be maintained for each vehicle model, and when an OEM wants to port an AI model
from one vehicle to another, the proper conversions can be handled
automatically via these mappings. Historically, maintaining these mappings has
been labour intensive, but LLM based systems can now automate much of the
translation and maintenance. This allows OEMs to amortise model development and
calibration across more vehicles, reducing both engineering cost and time to
deployment.
From your perspective, which edge AI capabilities will
deliver real customer value first?
Outside of ADAS, driver and occupancy monitoring, and
autonomy features, we see virtual sensors gaining interest among OEMs because
they reduce BOM costs with some pass through cost savings for customers. Fleets
are already using similar virtual sensors to lower operational costs.
Diagnostics and anomaly detection also have strong potential to improve the
owner experience. Much of that work remains cloud based for now, but we will
see things increasingly moving in vehicle in the next 12 to 24 months.
Which AI enabled vehicle functions already generate real
value in production today beyond showcase demos?
ADAS features like Automatic Emergency Braking, Active
Cruise Control, and Lane Keep Assist represent the most mature and widely
deployed in vehicle AI ML models. These deliver tremendous safety and driver
assurance benefits at scale. Driver monitoring features, including the
detection of drowsiness and distraction, are also quite mature and face
upcoming regulations requiring their use for key regions.
What are the biggest prerequisites in architecture, safety and toolchains for scaling AI across vehicle
portfolios?
Architecture is critical. Standardised runtimes, data
interfaces, and pipelines, plus the abstraction of data signals and hardware,
are key for scaling. Consolidating compute and selecting processors with neural
compute capabilities enable greater capacity for AI features. Safety requires
separating safety critical and non critical workloads, and deploying models in
containerised environments to help OEMs control the consumption of compute and
memory resources. Toolchains must provide robust tools for model version
management, observability into model KPIs and on vehicle resource usage, and
deployment mechanisms that support model promotion, relegation, and rollback
across diverse vehicle platforms.
Which architectural principles are shaping next
generation vehicle platforms today?
OEMs continue to target fully software defined networks with
service oriented interfaces, richer diagnostics, and fewer but more capable
high performance computing units within zonal architectures. After encountering
challenges with wholesale, cross platform architecture changeovers, however, a
number of OEMs have shifted towards a more incremental approach to the
transition, choosing to evolve existing architectures rather than replacing
them outright.
Where do legacy concepts still slow down scalability and
long term upgradability?
The legacy approach of an ECU for every function, coupled
with software and AI model integration at a signal level without abstraction,
forces every AI feature to become a custom integration project filled with glue
code. Outdated data capture policies can also slow model development and re-training,
while legacy over the air release trains built for infrequent updates conflict
with the iterative nature of AI model improvement. These patterns create
friction along the entire AI lifecycle.
Partnerships are becoming essential across the automotive
computing stack. Which type of partnerships will matter most in the next three
years, strategic, technological, or regulatory, and what should OEMs prioritise
first?
Technological partnerships will be the most important
over the next three years, particularly for scaling AI models across vehicle
fleets. The AI model lifecycle spans vehicle subsystems, data science, machine
learning, silicon, software development, and vehicle integration, too many
domains to maintain fully in house at a competitive cost. The most important
priority for OEMs is to identify an AI model toolchain that can meet their
needs while standardising runtimes, interfaces, and pipelines. By abstracting
the data and hardware, OEMs can then effectively leverage strategic
partnerships with model vendors and Tier 1 suppliers to scale.
