HM3D

Headlines · 2

1. Oct 13
   WEDGE uses an agent's failure modes to generate targeted navigation training worlds

   A preprint submitted to arXiv on 14 October 2026 by Yue Zhang and colleagues introduces WEDGE, a framework for weakness-guided navigation learning in which a pretrained agent's failure modes guide the generation of training environments at each adaptation round. WEDGE has an LLM propose executable weakness features from contrasting successful and failed rollouts, validates them on a diagnostic set, selects the strongest validated weakness by success-rate gap, and maps it to the environmental factors it then varies to build environments in the difficulty regimes where the policy struggles. Across three navigation agents with different adaptation mechanisms, the authors report distinct weakness profiles and different beneficial training environments, with weakness-conditioned adaptation consistently beating matched random and hard-example alternatives on ProcTHOR, and gains for Uni-NaVid and Gemini also extending to HM3D after ProcTHOR adaptation.

   Hugging Face · Papers · 🔥 0
2. Oct 7
   SuperNav: an agentic navigation system for tasks and scenes

   SuperNav keeps a pretrained multimodal LLM focused on interpreting requests, understanding scenes and making decisions, while delegating motion execution to navigation tools — no navigation-specific fine-tuning of the model itself. It combines Navigation Skills, agent-oriented tools for physical interaction, and task-progress and context management, tied together by a unified visual-point interface that lets the model specify destinations directly in images and revise decisions from execution feedback. The authors report it outperforms four evaluated baselines on instance-level, multi-object and demand-driven tasks, with category-level HM3D evaluation and a real quadruped robot deployment.

   Hugging Face · Papers · 🔥 0

Experience and discussion from the community

Share my HM3D experienceAsk about HM3D