NeuraFrame Embodied Fleet

Curated round-trip learning across a fleet of robots. Robots learn locally, you decide what comes back and what goes out, and nothing lowers a safety minimum.

Embodied Fleet is the server you run to manage learning across many machines. A robot learns something useful on the floor; you decide whether it is worth sharing; and if it is, you push it to the rest of the fleet. The heavy model stays in simulation, the robots stay light, and the fleet gets smarter together.

The Fleet server is a new capability you run in your own cloud or on-prem. It provides the round trip, the curation, and the safety and provenance. You stay in control of what is shared and what is pushed.

The round trip

Three stages, delta-based so it is cheap on bandwidth and storage.

robots  --collect deltas-->  Fleet server  --review + approve-->  you
   ^                                                                  |
   +-----------------------  distribute (signed)  --------------------+

Collect. Robots send only what is new since the last sync, and only what you flagged shareable.
Curate. You review the queue, clean and dedupe, and approve what should spread.
Distribute. Approved learnings go back to the fleet, signed, and merge into each robot.

Private stays home

Every learning has a scope. Device-private by default (one robot's quirk stays on that robot and never leaves), or flagged shareable (a skill worth spreading). Only shareable learnings are collected, and nothing is distributed without passing your review.

The merge law

Non-destructive. If a robot already knows a situation, a push does not overwrite it.
Corrections override, with provenance. A push you approve as a correction changes behavior on purpose, and it records who approved it and why.
Safety tightens only. A pushed constraint can raise a safety minimum but can never loosen a safety law, enforced on the device, so even a compromised server cannot weaken a robot's safety floor.

Ask the fleet

A robot that hits something new can ask the fleet before escalating to a human. If a peer already learned it, the robot gets the verified action and remembers it. If the fleet does not know either, it escalates to a human, and that answer can flow back up. Knowledge composes across the fleet, still with no model on any robot.

You decide the sync

Not every robot is always online, and not every deployment wants to spend memory. You set the policy: cadence (real time, daily, weekly, monthly, manual, or off), direction (collect only, distribute only, or both), scope and budgets, and manual triggers. Off is valid: an air-gapped robot simply never phones home.

Security and provisioning

On first run the Fleet server mints a fleet identity. Every push is signed, so a robot only accepts learnings from your fleet. On top of the transport TLS, you can set a passphrase vault so payloads are ciphertext to everyone but you. The fleet identity and vault passphrase are baked into your company-specific device package with the Fleet tool, so a robot is provisioned to talk to exactly your fleet. Rotating the identity requires re-provisioning devices, which is how you revoke.

Embodied is a new capability, and so is Fleet. The safety rules are a tool, not a guarantee, and you are responsible for what you approve and push, your own model and integration, and independent safeguards. See the EULA.

Operating the fleet

On the Fleet server host, install the server and run the installer as root. It mints your fleet identity, starts the service, and prints the identity to bake into your company-specific robot package:

sudo bash install_fleet.sh          # install the server, start it, print your fleet identity
nf-fleet provision                  # reprint that identity at any time
nf-fleet status                     # queue, approved, and robot counts

As robots report what they learned on the floor, curate it. The fleet only learns what you approve, and a pushed constraint can only tighten a safety minimum, never loosen one:

nf-fleet review                     # list learnings awaiting approval, with IDs
nf-fleet approve --all              # approve everything pending
nf-fleet approve 70e5 cfdd          # or approve specific learnings by ID
nf-fleet reject 9b21                # reject one you do not want to spread
nf-fleet policy set cadence weekly  # set the fleet sync cadence

Each robot runs the fleet client from the Embodied package, provisioned with your fleet identity (and vault passphrase, if set) baked in with the Fleet tool. It collects shareable learnings, asks the fleet before escalating to a human, and merges approved pushes on your cadence. Robots never run nf-fleet; that command is for the server only. Every command is in the CLI reference.

Get it

The fleet client ships in the Embodied package, so robots can join a fleet. Run the Fleet server yourself: download the Fleet server.

← Embodied Studio Fleet →