Research map / Multi-agent systems and games / Human-AI and goal-conditioned
THDA: Treasure Hunt Data Augmentation for Semantic Navigation
Oleksandr Maksymets, Vincent Cartillier, Aaron Gokaslan, Erik Wijmans, Wojciech Galuba, Stefan Lee, Dhruv Batra
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace RGB-D input with depth plus semantic segmentation and goal segmentation | RGB-D ObjectNav agent · ObjectNav val | Success (%): 8.0 -> 10.2 | Improved |
| add exploration reward to sparse success reward (ablation) | Sparse Success · ObjectNav val | Success (%): 0.0 -> 10.8 | Improved |
| use ExploreTillSeen reward instead of dense path-efficiency reward (ablation) | Depth +SemSeg +GoalSeg · ObjectNav val | Success (%): 10.2 -> 14.5 | Improved |
| use ExploreTillSeen reward instead of dense path-efficiency reward (ablation) | Depth +SemSeg +GoalSeg · ObjectNav val | SPL (%): 3.3 -> 4.8 | Improved |
| use sparse success-only reward instead of dense reward (ablation) | Depth +SemSeg +GoalSeg · ObjectNav val | Success (%): 10.2 -> 0.0 | Hurt |
| remove SemSeg, keep depth and GoalSeg (ablation) | ExploreTillSeen agent · ObjectNav val | Success (%): 14.5 -> 20.0 | Improved |
| remove SemSeg, keep depth and GoalSeg (ablation) | ExploreTillSeen agent · ObjectNav val | SPL (%): 4.8 -> 6.5 | Improved |
| replace predicted goal segmentation with ground-truth segmentation | ExploreTillSeen Depth + GoalSeg · ObjectNav val | Success (%): 20.0 -> 40.7 | Improved |
| add THDA pretraining then fine-tune on ObjectNav (ablation) | Depth + GoalSeg · ObjectNav val | Success (%): 20.0 -> 28.4 | Improved |
| add THDA pretraining then fine-tune on ObjectNav (ablation) | Depth + GoalSeg · ObjectNav val | SPL (%): 6.5 -> 11.0 | Improved |
| THDA pretraining only, zero-shot without ObjectNav training | Depth + GoalSeg · ObjectNav val | Success (%): 11.4 | Mixed |
| Depth + GoalSeg + THDA vs previous best Habitat leaderboard entry | SemExp / SRCB-robot-sudoer · ObjectNav test-standard | Success (%): 17.9 -> 27.4 | Improved |
Closest papers by meaning
- Auxiliary Tasks and Exploration Enable ObjectGoal Navigation (ICCV 2021 · similarity 0.89)
- PONI: Potential Functions for ObjectGoal Navigation With Interaction-Free Learning (CVPR 2022 · similarity 0.88)
- Scaling Data Generation in Vision-and-Language Navigation (ICCV 2023 · similarity 0.87)
- Bridging the Visual Semantic Gap in VLN via Semantically Richer Instructions (ECCV 2022 · similarity 0.86)
- Habitat Synthetic Scenes Dataset (HSSD-200): An Analysis of 3D Scene Scale and Realism Tradeoffs for ObjectGoal Navigation (CVPR 2024 · similarity 0.85)
- EfficientNav: Towards On-Device Object-Goal Navigation with Navigation Map Caching and Retrieval (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
