CHAIN of Consequences: A 3D Interactive Benchmark for Constraint-Aware Physical Planning

Physical intelligence requires more than recognizing the current scene or predicting isolated outcomes: agents must choose actions whose physical consequences determine what can be done next. We study this Level-III capability as constraint-aware sequential planning —anticipating how each action reshapes future feasibility. Existing evaluations largely target physical perception, physical prediction, or end-to-end embodied execution, leaving this capability poorly isolated. To address this gap, we introduce $\text{C}$ausal $\text{H}$ierarchy of $\text{A}$ctions and $\text{In}$teractions($\text{CHAIN}$), an interactive 3D, physics-driven benchmark for isolating Level III constraint-aware sequential planning. $\text{CHAIN}$ uses a fixed action API and closed-loop observations so that models must act, observe physical consequences, and update which future actions remain feasible. Its 200 task instances span five geometry-rich task families organized by three action–state coupling dimensions: constraint gating, spatial commitment, and structural propagation. Under a unified protocol, we evaluate state-of-the-art models, including VLM planners. Even the strongest model leaves more than half of the task instances unsolved, with failures clustering around early irreversible commitments, brittle replanning after state changes, and weak tracking of how constraints propagate through space and time.