Ground-Compose-Reinforce: Tasking Reinforcement Learning Agents through Formal Language

Grounding language in complex perception (e.g. pixels) and action is a key challenge when building situated agents that can interact with humans via language. In past works, this is often solved via manual design of the language grounding or by curating massive datasets relating language to elements of the environment. We propose Ground-Compose-Reinforce, a neurosymbolic framework for grounding formal language from data, and eliciting behaviours by directly tasking RL agents through this language. By virtue of data-driven learning, our framework avoids the manual design of domain-specific elements like reward functions or symbol detectors. By virtue of compositional formal language semantics, our framework achieves data-efficient grounding and generalization to arbitrary language compositions. Experiments on an image-based gridworld and a MuJoCo robotics domain show that our approach reliably maps formal language instructions to behaviours with limited data while end-to-end, data-driven approaches fail.