Research map / 3D vision / 3D generation
NextBestPath: Efficient 3D Mapping of Unseen Environments
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces AiMDoom, a Doom-based indoor active-mapping benchmark with four difficulty levels, and NBP, which predicts long-term goal value maps and obstacle maps to plan next-best paths instead of greedy next-best views.
Problem
Active 3D mapping of unseen indoor scenes requires efficient exploration trajectories, but greedy next-best-view methods get stuck in local areas and existing indoor datasets have limited geometric complexity and imperfect ground-truth meshes.
Approach
Predict a value map of accumulated surface coverage gain over a large spatial range around the agent plus an obstacle map (including predicted unseen obstacles), pick the highest-value cell as a long-term goal, and plan a Dijkstra path to it; train iteratively with online data collection, shortest-subpath data augmentation, and curriculum plus multitask learning.
Key results
- On AiMDoom Simple, NBP reaches 0.879 final coverage and 0.692 AUCs versus 0.760 and 0.605 for FBE and 0.599 and 0.479 for MACARONS.
- On MP3D, NBP achieves 79.38% completion versus 73.15% for ANM, a 6.23-point gain.
- On AiMDoom Hard and Insane, NBP reaches 0.618 and 0.472 final coverage, still well below full reconstruction, and is the best of the compared methods on every level.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use long-term goal and next-best-path planning instead of greedy next-best-view selection | NBP on MP3D · MP3D | Comp. (%): 73.15 -> 79.38 | Helped |
| Use long-term goal and next-best-path planning instead of the strongest baseline | NBP on AiMDoom · AiMDoom Simple, Normal, Hard, and Insane (unseen) | Final Cov.: 0.760 -> 0.879; 0.565 -> 0.734; 0.425 -> 0.618; 0.330 -> 0.472 | Helped |
| Use an oracle obstacle map instead of the predicted obstacle map during inference (ablation) | NBP · AiMDoom Normal | Final Cov.: 0.734 -> 0.808 | Helped |
| Use an oracle obstacle map instead of the predicted obstacle map during inference (ablation) | NBP · AiMDoom Normal | AUCs: 0.526 -> 0.580 | Helped |
| Train value-map and obstacle-map prediction jointly rather than as separate models (ablation) | NBP · AiMDoom Normal | Final Cov.: 0.712 -> 0.734 | Helped |
| Train value-map and obstacle-map prediction jointly rather than as separate models (ablation) | NBP · AiMDoom Normal | AUCs: 0.501 -> 0.526 | Helped |
| Use a 20m × 20m input crop instead of a 40m × 40m crop (ablation) | NBP · AiMDoom Normal | Final Cov.: 0.734 -> 0.630 | Hurt |
| Use a 30m × 30m input crop instead of a 40m × 40m crop (ablation) | NBP · AiMDoom Normal | Final Cov.: 0.734 -> 0.691 | Hurt |
| Use a 50m × 50m input crop instead of a 40m × 40m crop (ablation) | NBP · AiMDoom Normal | Final Cov.: 0.734 -> 0.647 | Hurt |
| Update long-term goals at each step instead of after completing a path (ablation) | NBP · AiMDoom Normal | Final Cov.: 0.734 -> 0.432 | Hurt |
| Update long-term goals at each step instead of after completing a path (ablation) | NBP · AiMDoom Normal | AUCs: 0.526 -> 0.367 | Hurt |
Surprising findings
- Frequently updating the long-term goal at every step (new strategy) performs much worse (0.432 final coverage) than updating only after completing a path (0.734).
- Using an oracle obstacle map at inference improves final coverage only from 0.734 to 0.808, so the value-map prediction is the main bottleneck.
Limitations
- Performance remains limited on Hard and Insane levels, with failure cases reaching only about 65% coverage in very complex scenes.
- Could not compare against UPEN and ANM on AiMDoom because their DD-PPO navigation policies were too costly to retrain.
- Long-term goal selection is greedy about high-value areas and can overlook previously seen regions with higher value.
Experimental setup
Datasets: MP3D AiMDoom
Benchmarks: AiMDoom MP3D
Metrics: Final Coverage AUCs Comp. (%) Comp. (cm)
Baselines: Random FBE SCONE MACARONS OccAnt UPEN ANM
Builds on: MACARONS frontier-based exploration Attention U-Net Boltzmann exploration learnable multi-task uncertainty weighting
Closest papers by meaning
- ActiveGAMER: Active GAussian Mapping through Efficient Rendering (CVPR 2025 · similarity 0.86)
- Habitat Synthetic Scenes Dataset (HSSD-200): An Analysis of 3D Scene Scale and Realism Tradeoffs for ObjectGoal Navigation (CVPR 2024 · similarity 0.86)
- GenNBV: Generalizable Next-Best-View Policy for Active 3D Reconstruction (CVPR 2024 · similarity 0.85)
- InternScenes: A Large-scale Simulatable Indoor Scene Dataset with Realistic Layouts (NeurIPS 2025 · similarity 0.85)
- SCRREAM : SCan, Register, REnder And Map: A Framework for Annotating Accurate and Dense 3D Indoor Scenes with a Benchmark (NeurIPS 2024 · similarity 0.84)
- RoomTour3D: Geometry-Aware Video-Instruction Tuning for Embodied Navigation (CVPR 2025 · similarity 0.84)
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.