Machineforce for robots
Let the model plan the path. Let something else approve it.
A language model is good at turning "move the pallets from bay 4 to dock 2, avoid the wet floor" into a route. It is not something you want steering a forklift on its own. A Machineforce unit puts the model to work as a planner with a job, bounds and an owner: it proposes the path, the bounds check it, a person approves it until the unit has earned more, and the robot only runs a plan that was checked and signed.
- The task, in plain words. An operator, a work order or another system asks the unit for a route: where to go, what to carry, what to avoid.
- A proposed path. The unit reads the site map, the robot's limits and its procedure, and proposes waypoints, speeds and stops. It says why, and cites the procedure steps it used.
- Checked against the bounds. Outside the model, every plan is checked against the unit's bounds: keep-out zones, speed and payload limits, working hours, which robots and which areas it may plan for. A plan that breaks a bound is refused, however it was asked for.
- A person approves. Until the unit has earned autonomy, a person with authority approves each plan with a yes or a no. "No, never route through the loading bay during a delivery" becomes a rule the next plan reads.
- Signed, then sent. Only an approved plan is signed and handed to the robot's own controller, which keeps its safety systems. The robot runs the plan; it does not take instructions from the chat.
- On the record. Every request, plan, check, approval and change is written to the Fluxive Ledger, so you can show later who asked for what and why the robot went where it went.
Why a unit, not a prompt
The rules live outside the model.
Safety rules pasted into a prompt sit inside the same model a confusing or hostile message is talking to. A unit's bounds are enforced after the model answers, so a clever request cannot talk its way past a keep-out zone.
Why a unit, not a script
It learns your site.
Corrections from the people on the floor become versioned procedure steps with an owner and an approver, so the unit plans the way your team works without a retraining run.
Earned, not assumed
Autonomy with a clock.
A unit asks for autonomy only after approved plans spread over different days. A person grants it, it expires, and one wrong plan takes it back.
On your own network
Nothing has to leave the building.
Robot pilots can run on your own network: the model, the map and the plans stay on a server you control, and the robot is reached only from inside. The cloud is optional, not required.
Where it fits
Warehouse and yard moves, inspection rounds, delivery routes inside a hospital or a campus, and site robots that are told what to do by people who are not robot programmers. The same loop works for any machine that takes a plan: a drone survey, a cleaning route, a camera sweep.
Under the software sits the enforcement chip, the CEU: a separate processor that checks each action before it executes and holds even if the software is fooled. It is at FPGA prototype today; the software guide-rails above work without it.