Attacca: Goal-Directed Control under State Continuity for Long-Horizon Embodied Agents
Attacca improves long-horizon Minecraft agents to 39-47.5% clean success, up to 7x better on extended tasks.
Attacca trains visual goal-conditioned policies for long-horizon embodied control where each task starts from the state left by the previous one and the next target may be outside view. It samples class-compatible masked goal images from other worlds, adds a target-mask prediction head, and conditions behavior on Search, Approach, and Interact phases. In Minecraft, clean success is 39.0-47.5%, 1.7-2.4x the strongest baseline. On long-horizon tasks it reaches 54%, 30%, and 28% completion, up to a 7x improvement.
- Policies train on full search-to-interact trajectories with goals decoupled from the scene.
- A target-mask head and Search, Approach, Interact phase conditioning guide control.
- Clean success is 39.0-47.5%, 1.7-2.4x the strongest baseline.
- Long-horizon Minecraft completion reaches 54%, 30%, and 28%, up to 7x better.
From the original article
1 min read at huggingface.co
A central capability of embodied agents is to accomplish complex objectives through sequences of interdependent tasks. Yet existing visual goal-conditioned policies underlying these agents are typically evaluated on isolated interactions where the target is already visible, and thus do not capture the conditions that arise during continuous long-horizon task execution.
By Gyusik Seo, Jaehong Yoon · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.