Embodied Graph: A Relational World Model for Long-Horizon Tasks
This paper proposes Embodied Graph, a relational state abstraction, and the Relational World Model (RWM) that performs imagination over graph states rather than dense pixel observations. RWM combines a Task-Aware Embodied Graph Constructor that filters redundant details and maintains graph memory, with an Embodied Graph Recurrent State Space Model that learns node-wise latent dynamics and inter-node dependencies. The authors report experiments on Minecraft and Crafter showing effectiveness in embodied environments, noting that RWM is the only method that obtains diamonds within the given budget in Minecraft, and further demonstrate applicability to web agents. The work targets long-horizon planning where delayed feedback and error accumulation make accurate imagination difficult.
It introduces a graph-based world model that improves long-horizon planning in embodied and web environments, with concrete Minecraft and Crafter evaluations.
Comments & Questions (0)
No comments yet
Be the first to comment!