Why do standard embodied planners struggle with long-horizon tasks? Storing experience as loose text or raw trajectories makes it impossible to isolate true causal dependencies from irrelevant scene clutter (like fridge color)!
CausGround introduces two core innovations:
1️⃣Factorized World-State:
Decomposes visual inputs into explicit object attributes (door: closed) and inter-object relations (holding: bread), discarding superficial scene noise.
2️⃣Bidirectional Causal Reasoning:
•Backward from Goal: Identifies mandatory Preconditions(e.g., to retrieve bread inside a microwave, the doormust be opened first).
•Forward over Effects: Anticipates environmental state changes before acting, eliminating repetitive deadlocks and trial-and-error.