RoboDreamer is a novel approach for learning compositional world models for robot imagination. The method leverages the natural compositionality of language to parse instructions into lower-level primitives, which are then used to condition a set of models for video generation. This enables compositional generalization, allowing the model to synthesize videos for new combinations of objects and actions. RoboDreamer also supports multimodal inputs, such as goal images and sketches, enabling more precise task specification. The model is trained using a probabilistic approach that encourages compositional reasoning by decomposing video generation into individual components. This approach allows RoboDreamer to generate videos for unseen tasks and successfully execute them in simulation. Experimental results show that RoboDreamer outperforms monolithic baseline approaches in video generation and demonstrates strong alignment with tasks under multimodal instructions. The model is evaluated on real-world robotics datasets and shows promising results in robotic planning tasks. RoboDreamer's ability to generalize to unseen tasks and incorporate multimodal inputs makes it a significant advancement in robot imagination and planning. However, the model has limitations, including its reliance on single camera views and its performance on real-world images. Future research could focus on improving generalization and incorporating multi-camera information.RoboDreamer is a novel approach for learning compositional world models for robot imagination. The method leverages the natural compositionality of language to parse instructions into lower-level primitives, which are then used to condition a set of models for video generation. This enables compositional generalization, allowing the model to synthesize videos for new combinations of objects and actions. RoboDreamer also supports multimodal inputs, such as goal images and sketches, enabling more precise task specification. The model is trained using a probabilistic approach that encourages compositional reasoning by decomposing video generation into individual components. This approach allows RoboDreamer to generate videos for unseen tasks and successfully execute them in simulation. Experimental results show that RoboDreamer outperforms monolithic baseline approaches in video generation and demonstrates strong alignment with tasks under multimodal instructions. The model is evaluated on real-world robotics datasets and shows promising results in robotic planning tasks. RoboDreamer's ability to generalize to unseen tasks and incorporate multimodal inputs makes it a significant advancement in robot imagination and planning. However, the model has limitations, including its reliance on single camera views and its performance on real-world images. Future research could focus on improving generalization and incorporating multi-camera information.