SCOPE: a camera agent and its digital twin
HRI 2026 · peer-reviewed · DOISmall AI models that run pan-tilt-zoom cameras from plain-English requests, on local hardware. We test them in a Blender digital twin with a 536-task benchmark.
Nikolaj Hindsbo
AI Engineer at Armada, Seattle
I build AI systems that run on local hardware, out where the data is, instead of in a distant data center. Most of my work is making small models dependable: camera agents that take plain-English instructions, forecasting models, and the tests that show whether any of it is ready to use.
M.S. in AI Engineering, Carnegie Mellon (2024). B.S. in Engineering Mechanics, UW–Madison (2023).
Selected work
Teaching an open model to read web pages
CMU, 2024We fine-tuned an open-source vision model (LLaVA) and rebuilt its prompts so it reads web pages better. Predicting what a click does went from 10% to 77.9%.
Waste classification on the edge
CMU, 2024Three image models that sort trash into six kinds, tested one change at a time, then compressed to run live on a Raspberry Pi 5 with almost no loss in accuracy.
Soft robot legs that bend the right way
CMU, 2024Pipe-crawling soft robots grip best when their legs bend outward. We simulated 12 leg designs, tested printed ones, and found one that does it reliably.
SpacExxon: A spacecraft that refuels satellites: Design & Analysis
UW–Madison, 2023Senior design capstone: a design study for refueling up to ten satellites per launch.