A grasping policy trained entirely on randomly assembled shape primitives, which then transfers to real objects it has never seen — removing the in-domain data collection step.
