Physical AI fails in system-software ways: a model that will not fit, a deadline missed under load, an update that cannot be rolled out safely. We work on all three, and we treat them as one stack rather than three separate fields. Each layer below is labelled with the timescale it governs.
Cloud-Native Computing
Deploying, updating, and orchestrating machine software the way the cloud does.
Mixed-Criticality Systems
Enabling safety-critical and best-effort workloads to share a system with strict isolation.
On-Device AI
Running AI workloads inside the machine, on heterogeneous CPU/GPU/NPU SoCs.
The three research axes, placed at the timescale each one governs.
On-Device AI
Autonomous vehicles and service robots increasingly need to run large models, including LLMs, without a round trip to the cloud. The hardware is heterogeneous and the memory budget is small, so the question is not whether a model fits but how it should be split, placed, and pipelined across the processors that are actually available.
- Optimal DNN partitioning under memory and bandwidth constraints
- Pipeline parallelism across asymmetric IoT nodes
- Online re-planning with bounded partitioning overhead
- Measuring what heterogeneous SoCs actually deliver for LLM inference
Publications on this axis
- DNNPipe: Dynamic programming-based optimal DNN partitioning for pipelined inference on IoT networks. Journal of Systems Architecture, 2025. doi
- Misconceptions and Realities of On-Device AI on Heterogeneous SoCs. Journal of Information Industrial Engineering, 2026.
Mixed-Criticality Systems
Consolidation is the defining move of AI-defined vehicles and modern robots: braking, perception, and infotainment end up on the same high-performance computer. Sharing hardware means sharing failure modes, so the runtime has to know which tasks must never miss a deadline and hold that guarantee even when the machine is overloaded.
- Criticality-aware CPU scheduling and temporal isolation in the kernel
- Admission control with probabilistic timing guarantees
- Budget escalation that contains, rather than spreads, an overrun
- Mapping mixed-criticality services onto a mixed-criticality runtime
Publications on this axis
Cloud-Native Computing
Vehicles and robots are becoming deployment targets. Kubernetes is arriving in the car, containers and virtual machines are being managed side by side, and domains built on different middleware stacks have to talk to each other. We work on the orchestration layer that makes this practical without giving up the timing properties the layers beneath it depend on.
- Kubernetes orchestration policies that respect real-time constraints
- Unified management of containers and virtual machines with KubeVirt
- Secure single-entry access to isolated VMs without VPN overlays
- Protocol-agnostic, automated inter-domain communication in AI-defined vehicles
Publications on this axis
- MCKube: a holistic mixed-criticality orchestration system in Kubernetes. Journal of Systems Architecture, 2026. doi
- A Protocol-Agnostic and Automated Approach to Bi-directional Inter-Domain Communication in Software-Defined Vehicles. International Journal of Automotive Technology, 2026. doi
- Design and Implementation of a Cloud-Native Virtual Machine Management System Based on KubeVirt. Journal of Information Industrial Engineering, 2026.