Preprint describes constraint-guaranteed robot planning method

A preprint describes Meta-Ctrl, a constrained-decoding framework for language-model robot plans that encodes syntax at the token level and preconditions, goals and ordering at the action level. The authors report that the factorization reduces constrained-decoding memory from more than 107TB to under 2GB. They also report a tabletop-robot demonstration in which generated plans satisfied encoded preconditions and goals by construction, alongside benchmark results using a small open-weight model.