FAME: exploration planning method transfers across scenes, beating NBP by 6.6 points on reconstruction
jiqizhixin · x · 2026-11-12
FAME, from Tianjin University and Hefei University of Technology (reported as a NeurIPS 2026 Spotlight), tackles planning under partial observation in active reconstruction: multiple unseen continuations can stay plausible behind the same local map.
Pipeline: Conditional Flow Matching generates multiple long-horizon anchors (each a possible exploration pose) → an A\-based planner converts anchors into feasible paths → clustering groups similar paths into exploration modes → hierarchical selection picks a mode first, then a trajectory within it.
Key result: trained only on AiMDoom and tested directly on Matterport3D without fine-tuning, FAME reaches 68.45% reconstruction completion vs. NBP's 61.81% — a 6.64-point transfer gain.
More from Research
- Understanding Deep Learning textbook updates chapter on measuring model performance — SimonPrinceAI · 2026-11-12
- Teknium on Unify's memory research: Hermes shows both memorization and ~15% gain over finetuning — Teknium · 2026-11-12
- Robotics take: get intelligence first, then distill to 27B sparse MoE — ChongZzZhang · 2026-11-12
- Meta releases MetaEncoder: one bi-encoder for retrieval and decision-making via natural language — pmttyji · 2026-11-12
- Physics-informed learning by IITM professor slashes AI simulation compute costs — ravi_iitm · 2026-11-12
- InvisibleMentor: AI That Watches Your Screen Recording to Suggest Better Workflows — burkov · 2026-11-12