Generative Flow on Symbolic Abstractions for Embodiment Capability Discovery

This research proposes FOSA, a framework that discovers embodiment-specific capabilities by learning a grounded task generator. FOS, a Generative Flow Network, constructs symbolic task skeletons which are compiled into simulation environments. Low-level policy performance provides reward signals for abstract policies, enabling the discovery of diverse and executable tasks across heterogeneous platforms. The code and demonstration videos are available at the provided link.