AssembleBench

As part of research at HUD developed AssembleBench, a 14-task contact-rich assembly benchmark on the NIST ATB-1 taskboard (peg insertion, gear meshing, nut threading) for evaluating and training VLAs on the DROID platform.

The work also includes a 1,355-episode synthetic demonstration dataset and a Code-Gated DAgger (CG-DAgger) loop that uses scripted experts to recover from detected failures. Full writeup here.