Attacca: Goal-Directed Control under State Continuity for Long-Horizon Embodied Agents
Attacca trains agents for the messy handoff between tasks, where the next target may not be visible from the state the last task left behind.
The paper frames that continuity problem as missing from standard visual goal-conditioned policy evaluations. Its method uses search-to-interact trajectories, goal images drawn from different but class-compatible worlds, target-mask prediction, and phase conditioning for Search, Approach, and Interact. In Minecraft tests, it reports 39.0-47.5% clean success, 1.7-2.4x over the strongest baseline. On long-horizon tasks, it reports 54%, 30%, and 28% completion, with up to a 7x improvement. HF Daily Papers' note
The paper frames that continuity problem as missing from standard visual goal-conditioned policy evaluations. Its method uses search-to-interact trajectories, goal images drawn from different but class-compatible worlds, target-mask prediction, and phase conditioning for Search, Approach, and Interact. In Minecraft tests, it reports 39.0-47.5% clean success, 1.7-2.4x over the strongest baseline. On long-horizon tasks, it reports 54%, 30%, and 28% completion, with up to a 7x improvement. HF Daily Papers' note
score 4