archipelago

a decentralized capability-sharing layer for autonomous edge systems.

archipelago is decentralized learning for collaborative autonomous platforms, starting with drones. each node learns locally, shares what it learned with its neighbors, and validates what it receives before adopting it. no central model is required.

the idea

platforms in contested, communications-constrained environments can’t count on a link to a central model. the bet is to share learned capabilities, not just observations: one platform learns something useful, another tests it locally and adopts it without repeating the learning.

how it works

each node trains its own pytorch policy. at episode boundaries a node offers a model version to a neighbor over an authenticated peer link. the receiver checks architecture, tensor shapes, finite values, size, age and provenance before anything changes its behavior, then accepts, adapts or rejects. every decision is logged. links can be cut, and nodes keep acting on their own until they come back.

the stack is python and pytorch for the runtime, fastapi and timescaledb for telemetry, grafana and a react dashboard for watching it, all in a docker compose setup.

what the evidence shows

it’s a hackathon prototype, and the experiment reports are explicit about their limits. in a pre-registered study, the learned policy beats idle, random and untrained ones on unseen worlds, but not yet a hand-written rule-based explorer.

a live pilot with four real node processes found that the exchanging fleet beat isolated nodes by a wide margin, from a single training run each, so it’s a lead and not a result. its partition and crash-restart checks did not pass. next up: terrain, richer validation, and keeping capabilities alive as nodes drop out.

githubdevpost