Proposal Discussion: Home Anchors and Learned Geographic Routing for Large-Scale Meshes

[JonoPorter]

So I have been thinking about the challenge of scaling a LoRa mesh beyond local and regional networks while staying within the strict bandwidth limitations that make Meshtastic successful.

I understand that traditional routing protocols are generally considered too expensive for LoRa, and that managed flooding performs surprisingly well within the intended network sizes. My goal is not to replace flooding, but to explore whether there are lightweight techniques that could reduce unnecessary retransmissions as networks grow.

Home Anchor Concept

Each mobile node could optionally designate a fixed "home" node.

Instead of trying to locate a mobile node by flooding a large portion of the network, messages could be directed toward that node's home anchor first.

The home anchor would maintain only minimal information:

• Last known relay node
• Last known geographic region
• Timestamp of last contact

For example:

Mobile Node A periodically checks in with Home Node A.

When someone wants to reach Mobile Node A:

1. Route toward Home Node A.
2. Home Node A reports where Mobile Node A was last seen.
3. The search area is reduced dramatically.

This resembles a rendezvous point rather than a full routing system.

Learned Geographic Routing

Rather than maintaining destination-specific routing tables, nodes could maintain a very small amount of learned geographic information.

Example:

Region → Preferred Direction → Confidence Score → Age

Rather than knowing exactly how to reach every node, a router would simply learn:

• Traffic from Denver usually arrives from the north.
• Traffic from Pueblo usually arrives from the south.
• Traffic toward a particular region historically succeeds through a specific neighbor.

This differs from traditional routing because the node is not storing paths to destinations. It is only learning successful traffic flow patterns.

Terrain Learning

One benefit of this approach is that it can naturally learn around obstacles.

For example:

The shortest path to a destination may be west, but a mountain blocks direct communication.

Over time the mesh may learn:

"Traffic for that region succeeds when forwarded north first."

The routing knowledge becomes:

Region → Successful Direction

rather than

Destination → Exact Path

This could potentially allow the network to adapt to geography while using far less state than conventional routing protocols.

Failure Handling

I would expect any such system to remain subordinate to flooding rather than replacing it.

A possible strategy:

1. Attempt learned directional forwarding.
2. If delivery fails, try alternate learned directions.
3. If confidence is low, fall back to managed flooding.

This keeps flooding as the ultimate reliability mechanism while allowing the network to benefit from accumulated knowledge.

Questions

• Has something similar already been evaluated?
• Are there known papers or prior experiments that combine rendezvous/home nodes with lightweight geographic learning?
• Would the airtime required to maintain this information outweigh any routing benefits?
• Are there scenarios where this could reduce retransmissions without introducing significant protocol complexity?

I am interested in feedback, especially from anyone who has investigated geographic routing, delay-tolerant networking, or large-scale mesh behavior in low-bandwidth environments.