Closed-Loop Physical AI: close the gap between plan and plant

Name: Technology
Uploaded: 2026-10-28
Duration: 4 min 39 s
Description: Imubit's leaders and an industry analyst on the models behind closed-loop control, why determinism matters and where industrial AI is going.

What Closed-Loop Physical AI is, why correlation-based AI cannot close the gap between plan and plant, and how Imubit maps, models and controls the decisions that can.

Read the transcript

Your plan is a best intention. Your plant runs on the decisions people make every shift, every hour, every minute in between. Every move is a trade-off. And every aspect of it is handled by a different team, on its own model, with its own version of the truth.

And a different person makes that call every shift, with different information. Imubit is Closed-Loop Physical AI for the process industries. In refineries, chemical and polymer plants, cement kilns and mineral processing. Physical, because every move lands on a running unit.

Closed loop, because the decisions get executed, always inside the envelope your engineers set. It starts with the map. Every business runs on a plan, built on economic and operational assumptions. Then there is how the plant really runs.

Map looks across the whole facility and finds the gaps between the two. It breaks each gap down to its causes, and to the recurring decisions behind it that can be improved systematically. Next comes the model, the cause and effect behind the map. How each decision moves each outcome, against its objective and its constraints.

Your engineers can make their own models and create their own competitive advantage. Then control, the autonomous execution of the key recurring decisions that close those gaps. Given the same inputs, the controller gives the same move, to every person, every time. That is what lets your operators trust it.

And every gap you close is measured, continuously, against the plan. Imubit is the only commercial solution applying closed-loop reinforcement learning at scale in a production environment. It has over 100 closed-loop controllers deployed worldwide, seven-plus years of closed-loop engagement in production, and thirteen approved U.S. patents. Imubit.

Map, Model and Control your gaps to plan. Every minute in between.

Closed-Loop Physical AI is industrial AI that models cause and effect in a running process and executes the key recurring decisions in closed loop, always inside the envelope your engineers set, and your operators can take over at any time. Imubit is Closed-Loop Physical AI for the process industries: it maps the gaps between the plan and the plant, models the cause and effect behind each decision, and controls the decisions that close those gaps.

The loop runs from plant data to a causal model, from the model to a decision, from the decision to a setpoint, and from the setpoint to a measured outcome that feeds the next decision.

The gap between plan and plant

A plan is a best intention. It depends on recurring operating decisions, made every shift, day and week by different people with different information and experience. Feed changes, equipment wears and the market moves, and each person answers those changes in their own way, so results change with who is on shift and what they know.

In between, every unit runs for itself and the space between them is run by hand. Every aspect of each trade-off is handled by a different team, on its own model, with its own version of the truth. The gap traces back to recurring decisions made that way, shift after shift. In chemical, cement and mining plants the plan often lives in multiple places; Map pieces them together.

Why correlation-based AI cannot close it

Analytics and correlation-based models can predict and recommend. They describe what moved together in the past. They cannot tell you what a deliberate move will do on a running unit, and that is the question every operating decision asks.

Process units are nonlinear and change over time, so the best operating point moves with feed, equipment condition and economics. Closing the gap takes a model of cause and effect that lets you see the result of a change before you make it, and a controller that acts on it every minute.

What "Physical" means

Every move lands on a running unit, with mass and energy balances and hard safety limits. So the AI behind it has to be deterministic: given the same inputs, it gives the same move, to every person, every time. You cannot let a language model improvise with your plant.

Imubit writes setpoints through your existing APC and DCS, never around them. ARC Advisory Group calls this human on the loop: the AI calculates and executes, and the operator moves from manual controller to supervisor.

How it works: Map, Model, Control

Map traces the gaps between plan and plant to the recurring decisions behind them, and ties each one to the objectives, constraints, disturbances and people involved. Read about Map.
Model captures the cause and effect behind each decision. Process engineers build nonlinear, dynamic models from the unit's own history with DLPC® Builder, with first principles built in and nothing in a black box. Read about Model.
Control executes the key decisions in closed loop. Reinforcement learning trains a controller across the regimes the unit has run, off site; your engineers certify it, and it runs minute by minute inside their limits. Every gap you close is measured against the plan. Read about Control.

When it is a fit

A measurable gap between the plan and how the plant actually runs
The same manual adjustments made shift after shift
Enough operating history and a defined scope of control
Clear objectives, limits and owners for the decisions involved
A decision where steady execution matters more than a one-time insight

Proof from production

Imubit is the only commercial solution applying closed-loop reinforcement learning at scale in a production environment, with over 100 closed-loop controllers deployed worldwide, 7+ years of closed-loop engagement in production and 13 approved U.S. patents.

1 to 3% average yield improvement
15 to 30% reduction in energy consumption, where energy is the optimization target
$0.25/bbl margin improvement in refining
In mining, minerals and metals: 1 to 3% metal-recovery improvement, 5 to 10% grinding-energy savings and a 2 to 5% throughput increase

ARC Advisory Group positions Imubit as a Leader in the ARC MarketMap 2026: Autonomous Execution Agents, Continuous Process Edition.

Results vary with the application, the starting point and how long the loop is engaged. Customers include Oxbow, for SmartKiln® petcoke calcination, and Preem. See customer stories.

Works with your AI, too

Your teams can ask the plant model what a change would do from the AI assistants they already use, and get the same answer every time. The language model asks; the deterministic model answers; the controller makes the move, inside your limits. See it on the Model page.

Common questions

Questions about Closed-Loop Physical AI.

Let’s get started

What is Closed-Loop Physical AI?

How is it different from AI agents and language models?

Language models and agents interpret information and coordinate digital work. Closed-Loop Physical AI decides and executes moves on a physical process, so it has to be deterministic: given the same inputs, it gives the same move every time. The two work together. Your assistants can ask the model what a change would do, and the controller makes the move, inside the envelope your engineers set. The chat never touches the controls.

What does autonomous mean here?

A certified controller executes selected decisions around the clock without a person entering each move, inside the limits your engineers set and through your existing APC and DCS. Your operators can stay in the loop or take over at any time.

What is the difference between causality and correlation in process control?

Correlation shows what moved together in the past. A causal model estimates what a deliberate change will do, separating the effect of the move from feed swings, equipment wear and other disturbances. That is what lets you see the result of a change before you make it.

Does Imubit replace APC or the DCS?

No. Imubit runs on top of your APC and DCS and works the space between the units. It extends what they do by executing the key optimization decisions in closed loop.

How do our operators stay in control of closed loop?

Closed loop always runs inside the envelope your engineers set. The controller is certified before it runs, every move stays inside those limits, and your operators decide when it runs. They can take over at any time.

Do our engineers need to be data scientists?

No. With DLPC® Builder, process engineers build, evaluate and retrain models of their own units from the plant's history, without Python or a machine-learning background.

What production experience does Imubit have?

Does it work outside refining?

Yes. Imubit runs in refineries, chemical and polymer plants, cement kilns and mineral processing. Where the plan lives in multiple places, as it often does outside refining, Map pieces them together.

Is our data clean enough to start?

In almost every case, yes. Send the historian export as it is, with outages, turnarounds and gaps included; you do not need to filter or clean it first. Our engineers screen it in the platform for missing data, compression, frozen tags and outliers, remove the periods that should not train the model, and agree with your engineers which tags matter. What can limit a use case is data that carries too little information, such as heavily compressed history, setpoints that never moved or lab samples without timestamps. If we find that, we tell you, along with what would fix it.

How do we start?

Call us to map your gaps. Your team prepares and uploads the data package in one to two hours, results arrive within a few working days of approval, and no workshops are needed.

From Transcend 2026

Technology

4:39

Imubit's leaders and an industry analyst on the models behind closed-loop control, why determinism matters and where industrial AI is going.

Read the transcript

Combining abstract mathematics with real-world critical systems has been a big passion of mine. And unlike chatbots and things like that, when you're deploying AI in the real world, the foundations are oftentimes necessary. In the real world, as opposed to playing Go, trial and error over a system could be something that's super costly or dangerous or both. So what you can do is take a model-based approach, which means that you do the trial and error, instead of over the system, you do it over a model of the system.

So it works like this. You start by taking historical data. You learn a model of the system. Then you take that model.

You train a controller. It gains experience through reinforcement learning. All of this happens offline in the cloud. And then you take that controller and apply it to the real system.

This is kind of teeing up what we're seeing. I've heard lots of discussion about, are we falling behind? And the answer is, well, if you're not one of the pacesetters, yes, you are. The closer you get to control, the more rigorous and more math and engineering-based you need to be.

I'd put Imubit firmly in this category that we're calling the autonomous execution agents, where they are using typically anything from neural nets, physics-informed neural networks, and various techniques. And they are more about human on the loop, or closing the loop, because they've got that rigor. So, kind of to clarify, our models are, the DLPC models are not large language models. They're actually deterministic once you train them as a deep learning network.

What it essentially means is that you're getting the same output for the same input, no matter which user and what time you ran this thing. And that's not how LLMs work. But certainly, my mental model states that if the problem is highly nonlinear, the gains are changing all the time, or if it's a truly multi-unit kind of optimization, then I think I would use an AI solution versus considering a legacy approach. We're not the orchestration layer.

And I think that's the point Gil was trying to make earlier, is that we were thinking, even very recently, we should build some agentic capabilities within our platform. And then what we did was we actually tested it ourselves, and we saw what Claude could do with our platform when it connected to it. And we're like, oh, we shouldn't do that. We should go this other route, where we actually enable LLMs, agentic capabilities, to meet you all where your journey is.

We're doing that because of the reason I mentioned, that we don't think we or anybody else or any other application platform is going to build one agent in a box and be the center of the universe. We want you guys to build your own enterprise architecture, with your own LLMs, your own workflows, and interface with us as an application platform. Because really, the LLMs are like a new web browser. That's really what it comes down to.

The biggest learning for me is the roadmap and the MCP that's coming up within the next few months. That's an exciting development, because now we can really start to unearth the power of data and models developed using other platforms and integrate that with the tools that Imubit has to offer. Everything that can be automated, that's instrumented and actuated in closed loop, is going to be, in the next few years, we believe, automated in closed loop. This example that I always use: you want the operator to be able to make the right decision as if they have all the cross-functional team behind them, but on a 2 a.

m. Sunday morning, where there's nobody behind them, nobody around them.

All the Transcend 2026 films

Where to start

Find out what your gap to plan is costing you.

One to two hours for your team to prepare and upload the data package
Results within a few working days of approval
No workshops

Send your data as it is. Gaps, outages and noisy tags are normal, and our engineers screen and clean them in the platform. If anything limits a use case, we tell you.

Let’s get started How a pilot works