Embodied agents must achieve complex objectives by chaining interdependent subtasks. Existing visual goal‑conditioned policies are typically benchmarked on isolated interactions where the goal is already visible, ignoring the state continuity inherent in long‑horizon execution. Each subtask starts from the state left by the previous one—different position, orientation, modified world, and the next target may lie outside the current field of view—so policies that rely solely on the current observation often fail to progress.\
Attacca addresses this by training visual goal‑conditioned policies on full search‑to‑interact trajectories. It uses context‑decoupled goal sampling, pairing each demonstration with a class‑compatible masked goal image from another world, thereby removing direct scene and pose correspondence. The network includes a target‑mask prediction head that provides dense grounding supervision beyond action imitation. Additionally, behavioral‑phase conditioning teaches the policy to distinguish Search, Approach, and Interact stages and adapt its control as execution proceeds.\
Evaluated on several short‑ and long‑horizon tasks in Minecraft, Attacca achieves 39.0%‑47.5% clean success on short horizons, a 1.7‑2.4× improvement over the strongest baseline. On long horizons it reaches 54%, 30%, and 28% completion, up to a 7× gain.\
Review