Latent Goal Prediction from Language for Model-Based Planning

The article introduces Latent Goal Prediction from Language (LAGO), a novel framework designed for model-based planning that predicts intermediate goal states from language instructions and action-conditioned rollouts within a shared latent space. LAGO addresses the limitations of traditional methods by dynamically decomposing instructions into tractable latent subgoals, allowing for coherent long-horizon planning without the degradation seen in prior approaches. This advancement is significant for practitioners as it combines the precision of visual goals with the flexibility of language, enhancing the effectiveness of planning in complex environments.