GSR: Learning Structured Reasoning for Embodied Manipulation

ABSTRACT

We introduce Grounded Scene-graph Reasoning (GSR), a structured reasoning paradigm that explicitly models world-state evolution as transitions over semantically grounded scene graphs. By reasoning step-wise over object states and spatial relations, rather than mapping perception to actions, GSR enables explicit reasoning about preconditions and consequences. We construct Manip-Cognition-1.6M, a large-scale dataset to supervise world understanding and action planning. Evaluations across RLBench and real-world tasks show that GSR significantly improves zero-shot generalization and long-horizon task completion over prompting-based baselines by treating scene graphs as the primary state space for decision making.