A hundred robots, one narrow passage

A robot at the front of an aisle stops. Someone has left a pallet in the way. The machine behind it keeps coming, while a third approaches from the other end. Each robot can recognise an obstacle. That ability alone does not tell the team who should wait, who should turn around or which delivery has now become late.

Now expand this imagined warehouse to a hundred machines. It is a useful picture of why physical AI reaches beyond putting a good model inside each robot. The job has become collective. The machines must share a changing place, limited communication and a plan that can no longer be followed exactly as written.

The obstruction changes more than one route

A longer route may be fine for the first robot and a problem for the rest. It could occupy a crossing another machine needs, delay a handover or send work toward a station that is already busy. Local avoidance handles one encounter; coordination has to consider what that response does to other tasks.

Open-RMF is a useful practical reference here. Its traffic schedules and negotiation mechanisms address how different fleets can share routes and building infrastructure. It makes the problem concrete: the team needs to exchange intended movement as well as present position, and update that account when plans change. ^[6]

Swarm intelligence is often used broadly for this area, although a coordinated fleet is not automatically a decentralised swarm. What interests me in either case is the same everyday outcome: does adding robots help finish more work, or merely create more waiting? A bigger team needs a way to make that difference visible.

The useful message is not always the largest one

For a robot approaching the pallet, “this passage has just become unavailable” may be more useful than a continuous video stream from every camera. Another robot doing an unrelated job may need none of it. Starting with the decision a message can change helps us choose what to send and who actually needs it.

MoRoCo studies how communication structure can adapt to different requests under limited links. A short status report, sustained video and remote operation need different arrangements. That framing is valuable: communication is part of the task, not a pipe whose capacity we assume will always be enough. ^[1]

The interactive model below compares everyone exchanging directly, neighbouring robots sharing and one link disappearing. Fewer lines are not automatically better. The question is whether the robot that needs the update receives it in time, and what it should do when it does not.

Figure 1

Who receives the news about the aisle?

Compare three communication arrangements. The lines show message paths, not measured traffic.

Choose a situation

Everyone exchangesNeighbour exchangeOne robot loses its links

1. Local viewWhat each robot observes
2. Useful exchangeChanges that matter to neighbours
3. Shared taskCoordination with visible limits

Every robot has direct links to the others.

What changed?

The example uses many direct communication paths.

In this example

An update can travel directly to every other robot.

What it means for the work

The design must still choose which updates deserve that reach.

The same team uses fewer direct links.

What changed?

Messages pass through neighbouring robots.

In this example

A relevant update needs a route to the robot that needs it.

What it means for the work

Less direct exchange is useful only if the job still gets the information.

Missing news is not evidence that the aisle is clear.

What changed?

A robot drops out of the neighbour network.

In this example

The remaining team must account for the missing update.

What it means for the work

Continuation depends on what the task can establish locally.

A shared picture can contain two different views

Return to the pallet. A floor-level camera sees an obstruction; an older overhead image appears to show a clear aisle. Averaging those descriptions into one confident status would hide the reason to look again. The team needs the observation time and viewpoint as well as the conclusion.

A high confidence score does not settle the disagreement by itself. Guo and colleagues’ calibration work shows why model confidence needs care before being read as a probability of correctness. For our example, keeping both observations available is more informative than letting the larger number erase the smaller one. ^[3]

Each robot also needs enough local capability to remain useful while the team catches up. Physical Intelligence’s knowledge-insulation research explores how action learning can preserve knowledge from a pretrained vision-language model. It is relevant to that capability, but does not remove the separate job of deciding what a neighbour’s old observation means now. ^[2]

Where quantum research might enter the plan

The pallet has made the next round of assignments harder. Some routes are unavailable, some batteries are low and some work can wait. Selected versions of such planning questions are where I would explore alternative solvers, including quantum methods. Prog-QAOA investigates more resource-efficient ways to represent optimisation in quantum circuits; it does not demonstrate a working advantage for this warehouse. ^[4]

The fair comparison counts the whole journey to a usable answer: preparing the problem, waiting for hardware, computing and checking. It keeps the same constraints for strong classical methods too. A solver that ignores a battery limit has answered a different question; a good route returned after dispatch time has missed its opportunity.

That leaves a sensible place for experimentation: planning that can tolerate the delay, with local movement and protection retaining independent classical controls. The team should still have a way to continue permitted work or wait if an experimental solver contributes nothing. The flow below shows that separation.

Figure 2

Keep shared information and permission distinct

Build a useful shared view

1. 01Qualify the observation

   Keep source, time and uncertainty.

   ↓
2. 02Share the useful change

   Send what another task depends on.

   ↓
3. 03Revisit the plan

   Account for missing or conflicting views.

Evaluate optional compute separately

1. 01Bound the question

   Choose a latency-tolerant planning task.

   ↓
2. 02Compare fairly

   Use the same constraints and full cost.

   ↓
3. 03Keep a classical path

   The local system must remain useful.

From a crowded aisle to a research programme

These are the questions we bring into our project-development work around CELTIC-NEXT and Horizon Europe: intelligence near the robot, coordination with constrained communication and fair evaluation of alternative computing. We are seeing strong interest. To make that useful, the next step is a shared problem that partners can test and a comparison whose outcome matters.

NODERIQ connects the themes as an applied research programme. Its starting point is a useful classical edge core; hybrid and quantum methods are extensions whose contribution must be shown. The international work remains at project and consortium development stage. A hypothetical busy aisle helps us describe the experiment we want, without pretending that the field deployment has already happened. ^[5]

This is why I think AI’s next major leap will happen in the physical world. It asks us to make learning, timing, communication and human work fit together. In our warehouse, success is not a persuasive explanation of the obstruction. It is a team that understands the change and gets useful work moving again.

Figure 3

The research behind the story

Further reading	The useful idea	The next question
MoRoCo · 2025 / 2026	The useful idea Studies team coordination and changing communication structure under restricted links.	The next question Which assumptions match the team and environment we actually have?
Knowledge Insulation · 2025	The useful idea Studies training a VLA action expert while preserving pretrained backbone knowledge.	The next question How does the resulting capability fit a constrained robotic system?
Prog-QAOA · 2025	The useful idea Investigates more resource-efficient quantum representations of optimisation problems.	The next question Does the complete planning path beat a strong classical reference?

Sources

1. MoRoCo: Multi-Robot Coordination under Restricted Communication arxiv.org
2. Physical Intelligence: VLAs that Train Fast, Run Fast, and Generalize Better pi.website
3. Guo et al.: On Calibration of Modern Neural Networks proceedings.mlr.press
4. Prog-QAOA: Resource-efficient Quantum Optimisation through Classical Programs quantum-journal.org
5. Neura Parse: NODERIQ Research Programme neuraparse.com
6. Open-RMF: Traffic Scheduling and Negotiation osrf.github.io