AI4Instruments: exploring laboratory measurement tasks in Isaac Lab

[Cherish0925]

Hi everyone, I'm Shuxing. I'm exploring laboratory research tasks for embodied AI through AI4Instruments, a proposal for AI participation in instrument design, development, and experimental use. My own work starts with optical experiments; I'd like to learn from people building robot learning environments in Isaac Lab.

One candidate task would have a robot position a probe or sample, select measurement locations and acquisition times, and use the observations to predict an unknown physical response. The agent could revise its measurement procedure as it learns. Later tasks could allow changes to the apparatus or sensor arrangement as well. This would give us a way to study how physical action, experimental design, and instrument development interact.

For a first evaluation, I would compare prediction error on previously unmeasured conditions at a fixed number of measurements or a fixed duration. An important design choice is which information reaches the agent: the readings an actual sensor could supply, their acquisition conditions, and the effects of calibration. The older sensor modelling discussion touches on some of these concerns.

Has anyone built an Isaac Lab task around active measurement, parameter identification, or a robot operating a scientific instrument? Which existing task would you start from, and what would you simplify first while preserving the experimental question?

AI4Instruments currently contains a paper, design notes, and a standalone optical simulation. The Isaac Lab task above is a proposal I would like to develop with input from this community. Pointers to existing work, concrete examples, and criticism of the task design would all be welcome.

Thanks,
Shuxing