TBNAI
Attacca: Goal-Directed Control under State Continuity for Long-Horizon Embodied Agents
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 05/10/2026, 20:00:00 UTC
Briefing created: 25/10/2026, 02:22:04 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 07/10/2026, 04:18:24 UTC.
Briefings
Attacca is a new method for long-horizon embodied agents that trains visual goal-conditioned policies on complete search-to-interact trajectories using goal images decoupled from the execution environment. The method uses context-decoupled goal sampling to pair demonstrations with class-compatible masked goal images from another world, learning dense current-view grounding through a target-mask prediction head. It introduces behavioral-phase conditioning to distinguish Search, Approach, and Interact stages, improving clean success rates by 1.7-2.4x on short tasks and up to 7x on long-horizon tasks in Minecraft.
Key points
- Attacca trains visual goal-conditioned policies on complete search-to-interact trajectories using goal images decoupled from the execution environment.
- The method achieves 39.0-47.5% clean success on short tasks and up to 7x improvement on long-horizon tasks in Minecraft.
Context
Embodied agents typically rely on visual goal-conditioned policies evaluated on isolated interactions where the target is already visible. This approach does not capture conditions arising during continuous long-horizon task execution, such as when the agent ends at a different position or the next target lies outside the current field of view. Attacca addresses this by training on complete trajectories with goal images from another world.
Limitations
The evaluation is limited to Minecraft environments for short- and long-horizon tasks. The method's success rates are measured on clean interactions, and the specific conditions under which the 7x improvement occurs on long-horizon tasks are not fully detailed in the provided evidence.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 25/10/2026, 02:22:04 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeibpcidvei5ovfxpfqnivj7p2f4wmkrhfvfhvh3sket2i4itbg5lhy