Attacca trains embodied agents for state-continuity, up to 7x on long-horizon tasks
nanyang-technological-university-singapore · hf · 2026-10-07
NTU researchers propose Attacca, tackling state continuity in long-horizon embodied tasks: each task starts from the state left by the previous one, and targets may fall outside the current field of view, breaking existing visual goal-conditioned policies.
- Method: trains on complete search-to-interact trajectories with context-decoupled goal sampling (masking goal images from other worlds to remove scene/pose correspondence), a target-mask prediction head for dense current-view grounding, and behavioral-phase conditioning distinguishing Search, Approach, and Interact stages.
- Results: on Minecraft, 39.0–47.5% clean success (1.7–2.4x over the strongest baseline); long-horizon completion of 54%/30%/28%, up to 7x improvement.
More from Embodied
- YC-backed Vexo launches always-listening AI bracelet that keeps your context — ycombinator · 2026-10-07
- H-JEPA: Hierarchical world models lift visual AntMaze success from 18% to 73% — ylecun · 2026-10-07
- DiVeR preprint trains VLA verifiers only on decision-critical states for test-time scaling — SharonYixuanLi · 2026-10-07
- Million-mile trucker after Tesla Semi factory tour: this industry is going to be changed — elonmusk · 2026-10-07
- WING transfers egocentric video skills to robots, hitting 99.2% on LIBERO — hongkongust · 2026-10-07
- Boston Dynamics Names Former Amazon Executive Rohit Prasad as CEO — SumitGup · 2026-10-07