NeurIPS paper: humanoids should imagine in 360° instead of thinking longer to find objects
_vztu · x · 2026-10-08
A NeurIPS 2026 paper, "Beyond Thinking: Imagining in 360° for Humanoid Visual Search," finds most frontier VLMs still struggle with the basic task of turning their head to find an object.
- Instead of long multi-turn Chain-of-Thought, the method asks models to imagine the semantic layout beyond the current field of view, then act on these spatial priors to decide where to look next.
- The idea enables a scalable data engine with 1.9M training samples and substantially improves active visual search across both specialized and frontier VLMs. Paper and code are available.
More from Embodied
- Hierarchy-GBP accelerates factor graph inference via coarse abstraction and recovery, SOTA BA runtimes — zhenjun_zhao · 2026-10-08
- EvoMem-VLA adds state-evolution memory for long-horizon robot manipulation, hitting 80.7% on RMBench — zhenjun_zhao · 2026-10-08
- EmbodiedSmith scales embodied data via recursive self-improvement flywheel in simulation — zhenjun_zhao · 2026-10-08
- Clear Oakley Meta Vanguard smart glasses earn Z87 jobsite safety certification — armand_ruiz · 2026-10-08
- DeepMind's Raia Hadsell on continual learning, navigation and real-world robotics opportunities at RAAIS — nathanbenaich · 2026-10-08
- Robot fights got too dangerous, so organizers put swords on the bots — cixliv · 2026-10-08