Centralized vs Decentralized

Collaborative (multi-robot) SLAM systems differ most fundamentally in where the global map is assembled.

Centralized architecture

A single server (ground station or cloud) aggregates data from all robots and performs global optimization. Each robot runs a lightweight local front-end (tracking, keyframe selection) and streams keyframes, descriptors, or submaps to the server; the server performs cross-robot place recognition, map merging, and global bundle adjustment, then broadcasts the corrected map back. Examples: C2TAM (cloud-based PTAM), CCM-SLAM (server + ORB-SLAM clients), maplab 2.0 (multi-session merging).

What flows over the network: keyframes (keypoints + descriptors + pose estimates) up; optimized map corrections down. Raw images stay on the robot.

A well-designed centralized system still degrades gracefully: CCM-SLAM clients keep tracking through dropouts by buffering keyframes and synchronizing retroactively, so the server is critical for the shared map but not for each robot’s safety.

Decentralized (peer-to-peer) architecture

There is no server: robots exchange information directly with neighbors when within communication range, and each robot maintains its own estimate of the (partially) global map. Optimization is distributed — each robot iterates on its own trajectory variables and exchanges only estimates of the boundary/separator poses with its neighbors, repeating until the network converges toward the global optimum. Examples: DOOR-SLAM, Kimera-Multi, Swarm-SLAM.

What flows over the network: compact place-recognition descriptors during encounters, then features for verifying candidate closures, then small per-iteration optimization messages — never whole maps.

Comparison

CentralizedDecentralized
Global optimizationat server, exactdistributed, iterative
Failure toleranceserver is criticaldegrades gracefully
Communicationrobot-to-serverrobot-to-robot (neighbors)
Outlier vettingserver sees all closuresrobust back-ends (PCM, GNC) required
Map versioningone authoritative maptemporarily divergent local views
Typical team sizesmall (2-4)larger swarms

How to choose

The fundamental tension is between consistency and communication cost: centralized systems achieve tighter global consistency but create a single point of failure; decentralized systems are more robust but harder to optimize globally. Recent work (Kimera-Multi, Swarm-SLAM) trends toward decentralized architectures with robust outlier rejection.

Why it matters for SLAM

The architecture choice drives everything else in a collaborative SLAM design — what gets transmitted, where loop closures are verified, how map merging happens, and how the system fails. For a small team with reliable WiFi, a centralized system is the simplest path to an accurate shared map; for swarms, subterranean exploration, or contested networks, decentralized designs are the only ones that survive.