Attacca: Goal-Directed Control under State Continuity for Long-Horizon Embodied Agents

Attacca. Goal-Directed Control under State Continuity for Long-Horizon Embodied Agents. By Gyusik Seo and one other, from Nanyang Technological University. Published on the sixth of October. Its code repository is worked out: about seventy source files, with a readme and a licence.

Embodied agents are meant to accomplish complex objectives through chains of tasks that depend on one another. The visual goal-conditioned policies underneath those agents, though, are typically evaluated on isolated interactions in which the target is already in view. That does not capture what happens during continuous, long-horizon execution. There, each task starts from whatever state the previous task left behind. The agent may be standing somewhere else, facing a different way. The world may have been changed. And the next thing it needs to interact with may be outside its current field of view.

A policy trained only on the easy case meets a hard one. If the agent cannot ground its target in what it currently sees, it may struggle even to begin the next task. The difficulty is not the interaction itself. It is getting from the end of one task to the start of the next.

Attacca trains visual goal-conditioned policies on complete trajectories that run from search all the way to interaction, using goal images that are decoupled from the environment the agent is acting in. The policy learns to find its target, not only to act on a target it has been placed in front of.

Three components carry this. The first is context-decoupled goal sampling. Each demonstration is paired with a masked goal image from another world, one whose class is compatible, so the goal no longer shares the scene or pose of the demonstration. The second is a target-mask prediction head that learns dense grounding in the current view, giving supervision beyond imitating actions. The third is behavioral-phase conditioning, which teaches the policy to tell apart three stages, Search, Approach, and Interact, and to adapt its control as execution moves through them.

Attacca is evaluated on several short-horizon and long-horizon embodied tasks in Minecraft. Clean success ranges from thirty-nine percent to forty-seven point five percent, an improvement of one point seven to two point four times over the strongest baseline.

On the long-horizon tasks, completion reaches fifty-four percent, thirty percent, and twenty-eight percent, with improvements of up to seven times. Those are the settings where state carries over from one task to the next, which is the condition the method was built for.

The gap Attacca addresses is the space between tasks. An evaluation where the target is always visible never asks a policy to search, and a goal image from the same scene hands it a direct correspondence that continuous execution will not provide. Decoupling the goal, grounding it in the current view, and naming the phases of execution are each aimed at the moment one task hands over to the next.