Communication constraints

Multi-robot SLAM lives or dies by its communication budget. Real robot teams talk over bandwidth-limited, intermittent links — WiFi at the edge of range, ad-hoc mesh radios, or underground/underwater channels measured in kilobits per second. Transmitting raw sensor streams or dense maps is infeasible, so collaborative SLAM systems are designed around what can be shared cheaply.

The cost hierarchy of shareable data

Roughly in increasing order of bytes-on-the-wire:

  1. Global place descriptor for a keyframe (a bag-of-words vector or learned embedding) — the cheapest currency; enough to detect a possible encounter.
  2. Local features of one keyframe (keypoints + descriptors) — enough to verify a candidate closure geometrically; orders of magnitude smaller than the image itself.
  3. Pose graph / sparse landmark subset — enough to align and jointly optimize trajectories.
  4. Compressed submaps (meshes, TSDF blocks, point-cloud segments) — the expensive tier, sent rarely, if at all.

Well-designed systems climb this ladder lazily: broadcast tier 1 continuously, send tier 2 only for promising matches, and reserve tiers 3-4 for confirmed merges.

Design patterns

Common pitfalls

A useful mental model: the communication channel is another sensor-budget constraint, like CPU or battery. Every architectural feature of collaborative SLAM — centralized vs decentralized, descriptor choice, submap granularity, loop-closure prioritization — can be read as an answer to “what fits through the pipe?”

Why it matters for SLAM

Algorithms that look identical on a benchmark differ by orders of magnitude in bytes-on-the-wire, and in the field the network is usually the first thing to fail. Designing for sparse, prioritized, dropout-tolerant sharing is what separates deployable multi-robot SLAM (subterranean exploration, search and rescue, warehouse fleets) from simulation-only results.