Research map / Multi-agent systems and games / Human-AI and goal-conditioned
MNE-SLAM: Multi-Agent Neural SLAM for Mobile Robots
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.
Proposes MNE-SLAM, a fully distributed multi-agent neural implicit SLAM system with peer-to-peer parameter exchange, intra-to-inter loop closure, and implicit submap fusion via distillation, plus a real-world indoor neural SLAM dataset (INS).
Problem
Neural implicit SLAM is limited to single agents and struggles in large indoor scenes and long sequences, while existing multi-agent frameworks are bandwidth-heavy and do not fuse implicit submaps consistently.
Approach
A fully distributed multi-agent neural SLAM where agents exchange only network parameters and RAFT features (not raw data) at loop closures, combining intra-agent and inter-agent loop closure with a rendering-based consistency loss and distillation-based fusion of implicit submaps.
Key results
- On Replica global tracking, RMSE 0.398 cm versus 0.559 cm for Point-SLAM and 0.453 for GO-SLAM (table 1 reports GO-SLAM at 0.404), and the method reports the best global reconstruction accuracy 1.681 cm.
- On ScanNet, average ATE RMSE 6.6 cm versus 9.7 cm for CP-SLAM and 6.9 cm for GO-SLAM across eight sequences.
- Communication is about 60.58 MB versus 429.78 MB for CP-SLAM, reported as roughly 7x more efficient.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add intra-loop closure (ablation) | MNE-SLAM · Replica | global reconstruction accuracy (cm) (↓): 1.831 -> 1.681 | Helped |
| add intra-loop closure (ablation) | MNE-SLAM · Replica | global tracking RMSE (cm) (↓): 0.432 -> 0.398 | Helped |
| add inter-loop closure (ablation) | MNE-SLAM · Replica | global reconstruction accuracy (cm) (↓): 1.903 -> 1.681 | Helped |
| add inter-loop closure (ablation) | MNE-SLAM · Replica | global tracking RMSE (cm) (↓): 0.441 -> 0.398 | Helped |
| add submap fusion (ablation) | MNE-SLAM · Replica | global reconstruction accuracy (cm) (↓): 1.949 -> 1.681 | Helped |
| add submap fusion (ablation) | MNE-SLAM · Replica | global tracking RMSE (cm) (↓): 0.401 -> 0.398 | Helped |
Surprising findings
- Single-agent tracking on Replica and ScanNet does not always reach the best; gains come mainly from multi-agent global tracking, as the authors note for Apartment.
- Most existing neural SLAM methods suffer significant pose drift on the INS real-world dataset and fail to reconstruct, per the authors.
Limitations
- Single-agent tracking on some benchmarks does not reach the best performance.
- Real-world INS results are reported for a limited set of methods, with several baselines failing.
Experimental setup
Datasets: Replica ScanNet Apartment INS
Benchmarks: Replica ScanNet Apartment INS
Metrics: Accuracy (cm) Completion (cm) Completion ratio (%) ATE RMSE (cm) runtime per frame communication (MB)
Baselines: iMAP NICE-SLAM Vox-Fusion ESLAM Co-SLAM GO-SLAM Point-SLAM PLGSLAM Loopy-SLAM CP-SLAM Swarm-SLAM CCM-SLAM
Builds on: NICE-SLAM DROID-SLAM Co-SLAM RAFT CP-SLAM
Closest papers by meaning
- MAGiC-SLAM: Multi-Agent Gaussian Globally Consistent SLAM (CVPR 2025 · similarity 0.90)
- Loopy-SLAM: Dense Neural SLAM with Loop Closures (CVPR 2024 · similarity 0.85)
- Query Quantized Neural SLAM (AAAI 2025 · similarity 0.85)
- DeepPointMap: Advancing LiDAR SLAM with Unified Neural Descriptors (AAAI 2024 · similarity 0.84)
- SNI-SLAM: Semantic Neural Implicit SLAM (CVPR 2024 · similarity 0.84)
- Learning to Navigate Efficiently and Precisely in Real Environments (CVPR 2024 · 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.