Decision software for robots
3Bird builds the decision layer for robots. It sits on top of the arm you already own, decides what to do with every item, acts when it is confident, asks a person when it is not, and records every decision.
Prefer to talk? Call +31 75 799 3038
Live decision feed
Threshold 85%
Illustrative example. Not live data.
The problem
A robot arm now costs less than a car. Yet most of them stop at the first item that does not match the script: a torn label, a new packaging, a product that arrived a day early.
Every new product and every new rule means another round of reprogramming. The integrator is booked, the line waits, and the business case slips. That is why so many lines with high variety and low volume are still done by hand.
Anything outside the programmed cases halts the line or needs a person to step in.
New products and new rules go back to the integrator for reprogramming.
When something goes wrong, nobody can show what the robot saw and why it acted.
The decision layer
A modern robot can see and it can move. What it lacks is judgment: what should happen to this item, right now, and is the answer safe to act on?
Fixed programs are fast, but they break on anything new. Language models can reason about exceptions, but they are slow, they write text, and they do not know how sure they are. The decision layer sits in between. It chooses only among the actions your rules allow, and it returns a probability that software can act on.
Your controller, your safety system and your integrator stay where they are. We add the layer above them.
Language model, only when rules or tasks change. Optional.
Seconds
Rules, decision model, confidence gate, audit log. This is what 3Bird builds.
About 1.5 s, while the arm moves
Camera, label reading and small vision models turn the scene into structured facts.
Fractions of a second
Robot controller, taught poses, hardware limits, guards and emergency stops.
Milliseconds
How it works
Four steps, on every item, every time. The same loop runs on any robot brand and with any decision model.
Camera and sensor data become structured facts about the scene: what the item is, its state, where it sits.
The model picks one of the actions the rules allow and attaches a confidence score to it.
Above the threshold the robot acts. Below it, an operator gets the case on a tablet and decides.
Every decision is logged with its confidence, the action taken and who approved it.
Each decision carries a score. You set the threshold per task. Above it the robot acts, below it a person is asked. Move the threshold and watch the outcomes change.
Illustrative example. In production the threshold is set per task, and a safety rule can block any action whatever the score.
Product demo
Two minutes in a simulated pharma cell. Routine boxes flow, a crushed box is blocked, an operator answers on a tablet, a rule changes in plain words, and every decision lands in the audit log.
The arm state, the current item with its facts, the candidate actions with their probabilities, and the decision feed.
The question to the operator: the reason and only the actions the rules allow. Options locked by a hard rule stay disabled.
A sentence becomes a structured rule. Priority, overlaps with existing rules and the assumptions are shown before confirmation.
One record per decision: timing, gate outcome, rules fired, model version and who approved.
Autonomy, escalations, confidence and decision latency, per run and over time.
Plain language rules
An operator types what should change. The decision layer turns it into a structured rule and waits for confirmation before it applies.
Operator console
Illustrative example
New rule
From now on, anything expiring before March goes to reject
Structured rule
| Condition | expiry_date < 1 March |
| Action | Reject |
| Scope | All items, all lines |
| Applies | From confirmation |
Awaiting confirmation
Edit Confirm
Technology
The decision layer runs on a new class of AI models. They do not write text. They choose between the actions the rules allow and return a calibrated probability for each.
Hard rules are deterministic code. The model can never override them.
It picks from the actions the rules allow. It never invents new ones.
Runs on a small box on your site. Images, decisions and logs stay in your network.
Any decision model, any robot brand, any camera. No lock-in to one vendor.
Above threshold
Robot controller
Runs the action
Below threshold
Operator console
One tap to decide on a tablet
Audit log
Every step, every decision, on the record
Daniel Kahneman describes two ways of thinking: fast, intuitive System 1 and slow, deliberate System 2. Most AI reasons like System 2, by writing text. Decision models work like System 1: they recognise a situation and choose.
| Language models | Decision models | |
|---|---|---|
| How it answers | Writes text word by word | Scores every option at once |
| Confidence | Not built in | Built in, per option |
| Actions | May invent new ones | Only allowed ones |
| Our use | Turning a sentence into a rule | Every live decision |
Use cases
The decision layer fits work where items change often and every exception costs time.
One handling step, for example loading vials into a tray, with your batches and your rejection rules.
One picking or sorting cell with your live mix of items.
One packing or loading station across at least two product variants.
One grading or packing line for one product family.
One service route or one handling task, with your staff answering the questions.
Where we are
The full decision loop runs today, on a local model with no cloud. We publish what we have measured and say what is still in progress.
0
wrong autonomous actions
200 hard test items
82%
of items handled without a person
the rest went to an operator
100%
of routine items decided correctly
same benchmark
0 ms
added wait for the arm (p95)
decisions are prepared while it moves
Benchmark on a synthetic set of hard cases in a simulated cell, September 2026. A strong start, not a safety claim: real items and real cells come next.
Pilot program
Four weeks, one cell, your items and your real exceptions. At the end you keep the full decision log, whatever you decide next.
Deliver cells that handle variety, and stop losing margin on every exception request.
Add a decision layer that meets on-premise data requirements and audit expectations.
Automate lines that were never worth a custom program.
Company
3Bird is a Dutch software company, founded in 2010 by Oscar Bout and based in Oostzaan, just north of Amsterdam. Our engineers work from the Netherlands and from Kathmandu, Nepal.
After years of building software for European businesses, we are putting that experience behind one problem: giving robots the judgment to handle the unexpected, while people keep control over the decisions and the data.
2010
Founded
100+
Software projects delivered
30+
Specialists
2
Locations: Oostzaan and Kathmandu
Give every robot the judgment to handle the unexpected: acting on its own when confident, involving people when needed, and making every decision transparent and auditable.
A world where people and robots work side by side with mutual trust, and where every business, large or small, can automate without giving up control over its decisions and its data.
FAQ
Short answers. Where something is not fixed yet, we say so and define it per pilot.
It is software that sits above the robot controller and decides what should happen to each item. It chooses among the actions your rules allow, attaches a confidence score, acts when the score is above your threshold, asks an operator when it is not, and records the decision.
Safety rules are deterministic code, and the model can only pick from the actions those rules allow. Below the threshold nothing moves until a person decides. Guards, light curtains and emergency stops stay with the cell. The decision layer adds judgment; it does not replace safety hardware.
None by default. The software runs on a box on your premises, and images, decisions and logs stay on your network. The only thing that can leave is the text of a new rule, if you use a hosted language model to write it. Remote support is off unless you switch it on.
Any robot whose controller accepts commands from an external system. The brands covered by a pilot are confirmed during scoping. ROS 2 support and more brands are on the roadmap.
It depends on camera access, controller access and the number of rules. Setup time is defined per pilot, and the pilot plan reserves its first days for it.
No. Your integrator keeps the cell, the safety concept and the robot program. The decision layer works alongside them, and we also work with integrators directly.
Pilot pricing is defined per pilot. Pricing after the pilot depends on the number of cells and is agreed before the pilot ends. There is no fee per decision.
Contact
Bring your items and your exceptions. We will show you what the robot does with them.
30 minutes, no obligation. We reply within one business day.
What do your robots handle, and where do they stop today? A person reads every message.
3Bird also supports clients with dedicated remote developers and AI-assisted software development, with European project management.
