Skip to main content
Back

University of Southern Denmark's Edge Cluster Makes Robot AI 30% Faster

A MAPE-K controller decides task-by-task whether AI inference runs on Raspberry Pis or Ryzen mini-PCs, cutting latency 25-30% and energy 15%.

Paper, Full Access
  • MAPE-K supervisor over K3s routes each LSTM inference to the best-fit pod using live latency, CPU, and power metrics.
  • Pod-level offloading cut average response time from 125 ms to 90 ms and tail latency from 300 ms to 180 ms.
  • Energy per inference dropped 15 percent by preferring busy Ryzen mini-PCs over saturated Raspberry Pis.
  • Custom Rust metrics server samples at 5 Hz instead of Kubernetes' 15 second default, enabling reactive decisions.
  • Deployed on a real assembly line with 5 humans, mobile robots, and Ubisense UWB tracking feeding a Keras LSTM.
  • Full paper on arXiv; no code release, agent may bottleneck at large scale.

Robots working alongside humans need to predict where those humans are about to walk, and they need that prediction in milliseconds. A team from the University of Southern Denmark tackled that constraint by building a Kubernetes-based edge cluster that watches its own response times and power draw, then reroutes each neural network inference to whichever node can serve it fastest. The paper reports 25 to 30 percent faster average response times and 15 percent better energy efficiency compared with a conventional service-level load balancer.

Why static offloading breaks in mixed human-robot spaces

Compute-intensive AI workloads in environments shared with humans have to run fast and cheap enough to stay on continuously. Cloud offloading helps with resource sharing and service availability, but it introduces security, latency, and adaptability problems across heterogeneous hardware. Pure edge deployments hit the opposite wall, where small devices saturate under load spikes.

The usual workaround is static or rule-based task allocation, which lacks flexibility for dynamic environments. Some approaches use heuristic or model-driven offloading, but they rarely integrate real-time feedback for AI tasks. That gap is what this work targets.

A MAPE-K loop wrapped around Kubernetes

The architecture uses the classic MAPE-K pattern (Monitor, Analyze, Plan, Execute over a shared Knowledge base) as an outer controller sitting above Kubernetes. The AI workload is an LSTM network that predicts human mobility, enabling proactive, conflict-free human-robot collaboration. Tasks are distributed across heterogeneous nodes (Raspberry Pi and ASUS PN53) managed via Kubernetes. A custom Rust-based metrics server provides real-time data on CPU load, latency, and power usage.