Speaker
Description
Vision-based navigation (VBN), based on passive optical sensing and AI methods, has become a key enabler for rendezvous, on-orbit servicing, and debris-removal missions. However, its validation continues to be challenged by a persistent domain gap between simulation and real-world operation.
Existing test strategies range from purely synthetic image-based evaluation, through camera- and hardware-in-the-loop approaches such as our Visual Servoing Testbed (VSTB), to robot-based facilities such as ESA’s GRALS, each with distinct trade-offs in realism, controllability, and cost. This talk compares three complementary testing paradigms for AI-based VBN systems relying on passive optical sensors: (1) pure computer simulation using synthetic imagery, for example with platforms such as DLVS3, which enables rapid iteration, broad scenario coverage, and low cost, but only limited representation of real sensor effects and hardware constraints; (2) our VSTB approach, a camera-in-the-loop, enclosure-based facility that recaptures rendered orbital imagery with a physical image sensor and space-relevant processing hardware to enable deterministic closed-loop validation while avoiding robot-induced artifacts; and (3) robot-based testing, where cameras observe physical mock-ups on multi-DOF robotic systems, as in GRALS, providing realistic scene geometry and accurate ground truth at the cost of more complex setups, fixtures, and limited trajectory coverage.
We discuss how these approaches address domain shift, sensor realism, real-time constraints, and hardware-in-the-loop requirements, and propose a layered validation methodology that combines synthetic, VSTB, and robotic testing as complementary steps toward future in-orbit demonstration.