Laboratory and Testbeds

Where the methods get tested before they reach an asset

Physical platforms and instrumented workflows, with what each one has actually been used for. Results from this equipment are reported in the papers listed under Q&A and Publications.

Physical platform

1:10 scale intelligent building-machine testbed

A scaled construction-machine platform in the structural engineering laboratory at Southeast University, commissioned in December 2025 and signed off against seventeen performance criteria. It couples a dual-arm collaborative robot and PLC control with a digital twin, so a construction sequence can be run physically and in simulation at the same time and the two compared. My group uses the platform; it was procured by the university, not by me.

17 criteria verifiedphysical + digital twin

Software stack

Layered robotics and AI development environment

A seven-layer research stack running from operating system and middleware through physics simulation and synthetic data generation to policy learning and on-board deployment, built on ROS 2 and NVIDIA Isaac Sim with domain randomisation and sim-to-real correction. It exists so that a control approach can fail in simulation a thousand times before it is allowed near a real structure.

ROS 2simulation to real

Instrumented workflow

Reality capture to structural assessment

Terrestrial laser scanning through point-cloud processing to a parametric BIM model and then to finite element analysis. Run end to end on a real bridge, this chain reached an Intersection over Union of up to 91% against critical structural outlines and identified 83% of primary components automatically. The numbers and the paper are on the Q&A page.

91% IoUpublished result

Instrumented workflow

Virtual trial assembly of prefabricated steel

Scan, model and analyse prefabricated components so the fit-up check happens before anything reaches site. On an experimental bridge the geometric prediction deviated by less than 1/1800 of the span, and predicted mean stress came within 90% of measured. Point-cloud registration accuracy, not the modelling, is what limits the whole chain.

<1/1800 spanpublished result
Every quantitative claim on this page is reported in a peer-reviewed paper with a DOI. The Q&A page gives each number as the paper states it, with the author position.