SwarmMap: Scaling Up Real-time Collaborative Visual SLAM at the Edge
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
SwarmMap is an edge framework that logs and replays SLAM operations to avoid sending full maps, uses a change-log synchronization layer, a priority-aware STS scheduler for optimization requests, and a map-backbone profiling (MBP) step that generates virtual keyframes to boost low-quality segments. Implementation is ROS-compatible and open-sourced. Evaluation and a three-month oil‑field pilot show at least 20% lower bandwidth, 95–105 ms end-to-end latency, 2× agent capacity (>20 agents), 38 cm average trajectory error and tracking loss under 4%.
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:34Background
- 3:03Scalability issues
- 6:11Key Insight
- 8:26Mapit - Map Information Tracker
- 10:31MBP - Map Backbone Profiling
- 11:15STS - SLAM-Specific Task Scheduling
- 12:30Experiments
- 16:04Conclusion
related talks
20mCarMap: Fast 3D Feature Map Updates for Automobiles
18mZooids: Building Blocks for Swarm User Interfaces
11mShape n’ Swarm: Hands-on, Shape-aware Generative Authoring with Swarm UI and LLMs
40mCo-operative robots sharing the load
43mPatterns for cloud-enabled robots
09mFusing Semantic Segmentation and Object Detection for Visual SLAM in Dynamic Scenes
29mIntegrate ROS into EdgeX - Startover of Edge Computing System
07mJan Erik Solem, Co-founder & CEO, Staer | Slush 2025