Research map / Multi-agent systems and games / Human-AI and goal-conditioned
MAGiC-SLAM: Multi-Agent Gaussian Globally Consistent SLAM
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 MAGiC-SLAM, a multi-agent RGB-D SLAM system that uses 3D Gaussian sub-maps, hybrid implicit tracking, DINOv2-based loop closure with pose-graph optimization, and coarse-to-fine map merging to produce a globally consistent, renderable map.
Problem
Multi-agent SLAM with novel view synthesis is slow, struggles to render real-world data, is limited to two agents, and has limited tracking accuracy and global consistency.
Approach
Replace distributed neural scene representations with rigidly transformable 3D Gaussian sub-maps, combine frame-to-frame registration initialization with frame-to-model re-rendering refinement for tracking, and add a centralized loop closure module that uses DINOv2 features for loop detection and pose-graph optimization over agents' sub-maps.
Key results
- On ReplicaMultiagent, MAGiC-SLAM achieves average ATE RMSE 0.27 cm (agent 1 avg 0.27) vs CP-SLAM 0.65 cm and Swarm-SLAM 1.42 cm.
- On training-view synthesis of ReplicaMultiagent, MAGiC-SLAM reaches 34.26 dB PSNR average vs 22.71 dB for CP-SLAM.
- Mapping takes 0.71 s per frame vs 16.95 s for CP-SLAM, and map merging takes 167 s vs 1448 s, with peak GPU 1.12 GiB vs 7.70 GiB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add loop closure (ablation) | MAGiC-SLAM · ReplicaMultiagent | ATE RMSE [cm]: 0.42 -> 0.27 (Off-0 average); 0.38 -> 0.16 (Apt-0 average); 0.54 -> 0.26 (Apt-1 average); 0.66 -> 0.32 (Apt-2 average) | Helped |
| add loop closure (ablation) | MAGiC-SLAM · AriaMultiagent | ATE RMSE [cm]: 2.13 -> 1.15 (Room0 average); 1.24 -> 0.65 (Room1 average) | Helped |
| add robust pose initialization (ablation) | MAGiC-SLAM · AriaMultiagent | ATE RMSE [cm]: 0.874 -> 0.670 | Helped |
| add robust pose initialization (ablation) | MAGiC-SLAM · ReplicaMultiagent | ATE RMSE [cm]: 0.825 -> 0.365 | Helped |
| replace NetVLAD with DINOv2 loop features (ablation) | MAGiC-SLAM · AriaMultiagent | ATE RMSE [cm]: 1.363 -> 0.900 | Helped |
| replace NetVLAD with DINOv2 loop features (ablation) | MAGiC-SLAM · ReplicaMultiagent | ATE RMSE [cm]: 0.351 -> 0.252 | Helped |
| use MAGiC-SLAM instead of CP-SLAM | multi-agent SLAM · ReplicaMultiagent | PSNR [dB]: 22.71 -> 34.26 | Helped |
| use MAGiC-SLAM instead of CP-SLAM | multi-agent SLAM · AriaMultiagent novel views | PSNR [dB]: 9.06 -> 22.61 | Helped |
| use MAGiC-SLAM instead of CP-SLAM | multi-agent SLAM · AriaMultiagent novel views | Depth L1 [cm]: 15.18 -> 3.15 | Helped |
| use Gaussian sub-maps and MAGiC-SLAM mapping | MAGiC-SLAM · ReplicaMultiagent office0 | Mapping frame time [s]: 16.95 -> 0.71 | Helped |
Surprising findings
- Gaussian-SLAM fails on AriaMultiagent and gives very large errors on ReplicaMultiagent (average 38.21 cm), indicating poor robustness in this setting.
Limitations
- Operates slightly faster than 1 FPS; implicit tracking requires many iterations to converge.
- Not designed for dynamic environments; real-world data had to be filtered to remove dynamic objects.
- Scalability constrained by the centralized server's capacity.
Experimental setup
Datasets: ReplicaMultiagent AriaMultiagent
Benchmarks: ReplicaMultiagent AriaMultiagent
Metrics: ATE RMSE PSNR SSIM LPIPS Depth L1 tracking frame time peak GPU memory
Baselines: CP-SLAM CCM-SLAM Swarm-SLAM Gaussian-SLAM MonoGS ORB-SLAM3
Builds on: 3D Gaussian Splatting Gaussian-SLAM MonoGS CP-SLAM DINOv2
Closest papers by meaning
- MNE-SLAM: Multi-Agent Neural SLAM for Mobile Robots (CVPR 2025 · similarity 0.90)
- GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting (CVPR 2024 · similarity 0.88)
- DG-SLAM: Robust Dynamic Gaussian Splatting SLAM with Hybrid Pose Optimization (NeurIPS 2024 · similarity 0.87)
- MAC-Ego3D: Multi-Agent Gaussian Consensus for Real-Time Collaborative Ego-Motion and Photorealistic 3D Reconstruction (CVPR 2025 · similarity 0.87)
- VTGaussian-SLAM: RGBD SLAM for Large Scale Scenes with Splatting View-Tied 3D Gaussians (ICML 2025 · similarity 0.87)
- Gaussian Splatting SLAM (CVPR 2024 · similarity 0.87)
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.