SPEAKER: Eduardo Kochenborger Duarte, Postdoc at ITU.
ABSTRACT: Testing robot manipulation systems under different real-world conditions can be challenging, as physical testing is often costly and time-consuming. This talk presents ongoing work on testing and uncertainty quantification for robot manipulation, using a robot arm and Danfoss components as a case study. The current focus is on investigating how synthetic data generated through CAD rendering, Gaussian splatting, and image mutations can be used for conformal calibration. In particular, we are interested in how closely synthetic data reproduces real-world observations, and how differences between synthetic and real data distributions affect the reliability of uncertainty estimates. The broader goal is to understand the possibilities and limitations of using synthetic data for testing robotic systems.