THANK YOU FOR YOUR ATTENTION

Trust, selective cognition and the reflexive intelligence robots need around humans

A motorcyclist travelling through traffic doesn’t care equally about everything in view. The colour of the building beside the road doesn’t matter. Neither does the make of the parked car, the pattern on a billboard or the precise shape of the clouds overhead.

What matters is the vehicle drifting towards the lane. The driver waiting to pull out. The pedestrian obscured by a van. The gap that’s closing faster than expected.

The working rule is blunt:

We only care about things that could kill you.

It isn’t literally true of course. Plenty of things deserve attention without being fatal. But as a compact principle for operating in a complex physical world, it’s useful. Pay attention to what could affect you. Give more attention to what could cause harm. Increase attention as uncertainty, proximity and possible consequence rise.

Humans do this without thinking. Robots don’t. Today’s robots can perceive extraordinary amounts of detail. They can identify people, objects, surfaces and free space. They can build geometric maps and generate plans towards a goal. But they still struggle to decide what matters now.

That problem sits beneath many of the awkward behaviours we see in Physical AI. Robots freeze in ambiguous situations. They react late to people. They leave excessive buffers. They repeatedly replan around movements a human would understand instinctively.

They can see the room. They just can’t read it.

This paper proposes that selective attention is a foundational requirement for human-aware robotics. Not attention in the narrow transformer sense, and not a visual saliency map. This is embodied attention: the continuous allocation of machine intelligence towards the people, motions and possible futures that matter most.

The central claim is simple:

Before people can trust robots, robots must demonstrate that they’re paying attention to us.

That requires more than detecting a person. It requires understanding how that person is moving, what their movement may mean, what could happen next, and whether the robot’s own actions could create harm, discomfort or surprise.

Attention is where motion becomes relevance. It’s where relevance becomes consequence. And it’s where machine behaviour begins to earn human trust.

Trust is the bottleneck

Physical AI is moving rapidly into human environments.

Humanoid robots are entering warehouses and factories. Mobile robots are operating in hospitals, airports and public spaces. Autonomous vehicles share roads with human drivers. Home robots are being designed to carry objects, prepare food and work within arm’s reach of families.

The technical question has long been whether these systems can perform useful tasks.

The commercial question is becoming whether people will trust them enough to buy, deploy and live beside them.

A family robot might be capable of chopping an onion perfectly. But would you reach past it for a bowl while it held the knife?

A warehouse robot might be able to carry a heavy load across the floor. But will workers trust it to understand when they’re about to step into its path?

An autonomous vehicle might obey every explicit road rule. But will passengers trust it when it hesitates, brakes unexpectedly or fails to recognise what another driver is about to do?

Capability is necessary. Trust determines adoption.

People won’t trust robots because the robots contain a certified perception model, a sophisticated planner or a long list of safety constraints. Most people will never inspect any of those systems.

They’ll judge the robot through its behaviour.

• Did it notice me?
• Did it leave enough room?
• Did it slow before I had to step away?
• Did it understand that I was reaching past it?
• Did it respond to my hesitation?
• Did it act before the situation became unsafe?

Human trust is built from repeated evidence that another agent is aware of us and capable of accounting for our likely behaviour. A robot earns trust in much the same way.

Trust isn’t something the machine declares. It’s a conclusion the human reaches.

Attention is the precursor to trust

A robot can’t account for a human it hasn’t noticed. It can’t anticipate behaviour it hasn’t interpreted. And it can’t interpret every agent, object and possible future with equal depth. The physical world contains too much information.

A robot operating in a busy train station might detect hundreds of people. A vehicle might perceive dozens of nearby agents and thousands of static scene elements. A humanoid in a home may encounter people, pets, furniture, appliances and objects whose relevance changes from one second to the next. Uniform cognition isn’t practical. Nor is it intelligent.

Humans don’t reason deeply about every observable element in a room. We maintain a compressed model of the situation and devote additional thought to the small number of things that might affect us. Most of the world stays in the background – a few things move into focus. Selective attention is the mechanism that controls that transition.

For Physical AI, it should answer:

• Who matters right now?
• What has changed?
• Which movement is ambiguous?
• Where might intent be emerging?
• Which possible future deserves prediction?
• What could happen if the robot continues its current action?
• How much machine intelligence should be allocated to finding out?

This leads to a broader interpretation of attention: Attention isn’t simply a weighting applied inside a neural network. It’s the allocation of cognition according to possible consequence.

Four systems of machine cognition

Human thinking is often divided into fast and slow forms of cognition. Robotics needs a slightly different structure because physical machines also contain a layer beneath cognition: direct mechanical execution.

A useful model is to divide robotic behaviour into four systems.

System 0: mechanical execution

A motor turns. A joint bends. A wheel changes speed. A gripper closes. There’s no interpretation here. The system has received a command and executes it within physical and control constraints.

System 0 contains the learned motor skills, control loops and stabilisation mechanisms required to act. A humanoid maintains balance. A mobile robot follows a trajectory. A robotic arm moves towards a target pose. It’s the body doing what it has been told to do. System 0 asks no questions about whether the action is socially appropriate, whether a human is about to intervene, or whether the plan remains safe as the world changes. It just moves.

System 1: instinct

System 1 is fast, continuous and reflexive. This is where the robot notices that a person is stepping towards its path. It recognises hesitation, drift, convergence, instability or a sudden change in movement. It elevates attention before the interaction becomes an explicit emergency. System 1 doesn’t solve the robot’s whole task. It doesn’t decide the long-term goal. It maintains awareness of what is changing and what could matter.

For humans, System 1 lets us move through a crowded room without calculating every trajectory. It helps us recognise that someone is about to cross in front of us, even before they’ve committed to the movement.

For robots, System 1 should provide the equivalent reflexive interpretation.

It asks:

• What’s moving?
• What does that movement suggest?
• Who needs attention?
• Is the situation changing?
• Do I need to slow, wait or create space?

This is the natural home of Motion Intelligence, intent inference, human awareness and consequence-weighted attention.

System 2: planning

Deliver the package. Carry the tray. Clean the room. Drive to the destination. Move the component from one station to another.

This is the domain most autonomy stacks already understand well. A planner receives a world representation, considers constraints and generates a sequence of actions towards the current objective.

System 2 is deliberate, but bounded. It doesn’t need to reconsider the nature of work, formulate a new life goal or engage in open-ended philosophical reasoning. It needs to complete the task safely and efficiently.

The critical architectural point is that System 2 shouldn’t be responsible for discovering every emerging human risk from first principles. It should receive a world state already shaped by System 1.

Planning should act on understanding. It shouldn’t have to reconstruct understanding every time it plans.

System 3: deep cognition

System 3 is sustained, abstract and open-ended reasoning. It includes deep reflection, long-horizon strategy, moral judgement and consideration of questions far beyond the immediate task. Humans use forms of this cognition when we rethink our goals, question assumptions or reason through complex unfamiliar problems.

Robots may not need a meaningful System 3. A robot carrying a tray doesn’t need to contemplate the social history of hospitality. A warehouse robot doesn’t need an inner philosophy of labour. A vehicle doesn’t need to reflect on the nature of travel.

General-purpose cognitive models may still support unusual planning problems, explanation, instruction-following or human communication. But deep cognition shouldn’t be confused with the reflexive intelligence required for safe physical interaction.

• A robot can be highly articulate and still fail to notice a child entering its path.
• It can reason through a complex written problem while remaining socially clumsy in a hallway.

Physical trust will be determined much more often by System 1 than by System 3.
The urgent challenge for robotics isn’t making machines think more deeply about everything. It’s helping them notice the right things quickly enough.

State is all you need

Selective attention requires a working representation of the world. Not a full recreation of every pixel. Not an exhaustive simulation of every possible future. Not a static map that becomes stale as soon as someone moves.

The robot needs a compact, continuously updated account of what matters. We call this the World State Vector.

The World State Vector isn’t the world itself. It’s the machine’s current understanding of the world, compressed around relevance.

It may contain:

• active agents
• relative position and movement
• motion-token history
• inferred intent
• behavioural trajectory
• confidence and uncertainty
• proximity
• trajectory convergence
• novelty
• prediction error
• possible consequence
• current attention level
• relationships between agents
• relationships between each agent and the robot’s plan

The World State Vector answers three immediate questions:

• Who is here?
• What are they doing?
• What are they likely to do next?

But its deeper role is architectural: it acts as the interface between perception, instinct and planning.

Perception updates the state.

Motion Intelligence interprets change within the state.

Selective attention decides which parts of the state deserve deeper computation.

Consequence modelling projects the most relevant parts forward.

Planning receives the resulting understanding and decides what to do.

The architecture becomes:

Perception → World State Vector → selective cognition → planning → System 0 execution

The World State Vector is persistent, but the intelligence acting on it can be dynamic.

A parent model and temporary child intelligences

A robot in a crowded environment shouldn’t run a maximum-depth model for every visible agent at all times.

Most agents and most situations won’t affect the robot. Some require light monitoring. A small number deserve active prediction. Perhaps only one or two require immediate consequence simulation.

This suggests a selective machine-learning architecture made up of one parent model and a changing population of child models.

The parent maintains the World State Vector. It watches for changes in:

• motion
• proximity
• trajectory convergence
• uncertainty
• novelty
• prediction error
• interaction risk
• relevance to the current plan

When an agent becomes important, the parent allocates more cognition to it.

A lightweight Motion Intelligence child process is created for that agent. If the situation becomes more consequential, the child can deepen its analysis, expand the prediction horizon or branch into multiple possible futures.

When the agent moves away, stabilises or ceases to matter, the child model can be reduced or removed.

These child intelligences come into and out of existence as the scene changes. The architecture isn’t trying to understand everything equally. It’s trying to understand the right things at the right depth.

A simple attention ladder might be:

Ignore

The agent has no plausible interaction with the robot or its current task.
No dedicated child model is required.

Monitor

The agent is present and potentially relevant, but its behaviour is stable.
The parent retains a low-cost state representation.

Track

The agent’s motion, proximity or relationship to the robot requires persistent observation.
A lightweight child model maintains a detailed motion history.

Predict

The agent’s intent is ambiguous, changing or convergent with the robot’s path.
The child predicts likely future motion-token sequences and updates them continuously.

Simulate consequence

The agent’s possible futures could materially affect safety, comfort, task success or human trust.
The architecture evaluates what may happen under different robot actions.

The child models aren’t necessarily separate large neural networks loaded from scratch. They may be dynamic inference contexts, specialised heads, small recurrent processes or selectively activated pathways through a shared model.

The implementation can vary, but the principle remains:

Machine cognition should scale with relevance and consequence.

Motion tells the system where to look

The architecture still needs a reliable early signal. Motion is the most universal candidate.

Humans communicate intent through movement continuously, often without realising it.

• A slight deceleration can indicate yielding.
• A lateral drift may precede a lane change.
• A pause can signal uncertainty.
• A change in posture can show commitment to movement.
• A narrowing gap can turn two independent trajectories into an emerging interaction.

Individually, these are small physical changes. Over time, they form the grammar of movement.

SpatioTemporal’s Motion Intelligence approach compresses these changes into discrete motion tokens. Short overlapping windows of movement are represented through kinematic properties such as bearing, distance, velocity, acceleration and jerk.

The result isn’t an image of the person or vehicle – it’s a compact description of how their movement is evolving.

Transitions between these tokens can reveal patterns such as:

• steady approach
• controlled yielding
• unstable drift
• hesitation
• acceleration
• divergence
• convergence
• sudden interruption
• repeated correction
• commitment to a path

These signals are useful because they often emerge before an overt action is complete. By the time a pedestrian is fully inside the robot’s path, perception can detect the hazard easily. The more valuable question is whether the robot could recognise the beginning of that decision earlier.

Motion Intelligence gives selective attention something to act on.

A change in the motion-token sequence may increase uncertainty.

• Unexpected acceleration may increase consequence.
• Converging trajectories may increase relevance.
• Prediction error may indicate that the existing model of the person is no longer sufficient.

Motion isn’t the entirety of human understanding. Semantic context, gestures, gaze, signage and social convention will still matter. But movement is the continuous physical signal available across almost every shared-space interaction. It’s where intent first becomes visible.

Consequence-weighted attention

Attention shouldn’t be determined by proximity alone. The nearest agent isn’t always the most important: a person standing beside a stationary robot may be closer than a child running towards its future path. A vehicle several car lengths ahead may matter more than the car immediately beside the observer if its unstable motion could trigger a chain of braking.

Nor should attention be determined only by likelihood. A very probable but harmless event may deserve little computation. A less likely event with severe consequences may deserve immediate focus. The useful quantity is expected consequence.

For agent i, a conceptual attention score could be expressed as: A_i = f(R_i, U_i, N_i, P_i, C_i, G_i) where:

• A_i is the attention allocated to agent i
• R_i is interaction relevance
• U_i is uncertainty
• N_i is novelty or deviation from expected behaviour
• P_i is the probability of a consequential interaction
• C_i is the severity of that consequence
• G_i is relevance to the robot’s current goal

A simplified form is:

A_i \propto P(C_i \mid M_i, W, G) \times I(C_i)

where:

• M_i is the agent’s observed motion
• W is the current World State Vector
• G is the robot’s current goal
• P(C_i) is the probability of a consequential interaction
• I(C_i) is the impact of that consequence

This isn’t proposed as a final universal equation – it expresses an architectural priority:

Attention should rise when uncertainty and possible consequence rise together.

This is the more precise version of “we only care about things that could kill you”.

For robotics, the principle becomes:

We care most about things that could cause harm, create conflict, violate expectations or destroy trust.

The same architecture should care about a robot cutting someone off, startling them, invading their space, blocking their movement or behaving in a way that appears unaware. Trust can be damaged long before physical contact occurs.

Safety isn’t enough

A robot can avoid collisions and still be difficult to trust.

• It might brake at the last possible moment.
• It might move too close before correcting.
• It might freeze unpredictably in a doorway.
• It might continue towards someone until a hard safety threshold is crossed.
• It might technically yield, but do so late enough that the human can’t tell whether they’ve been seen.
• These behaviours may satisfy a narrow safety condition while still feeling unsafe.

Humans care about legibility. We want to see that another agent has noticed us and adjusted accordingly. This is why early yielding matters.

When a driver slows with enough distance, a pedestrian knows they’ve been seen. When a person creates space before another person reaches them, the interaction feels fluent. When someone pauses to let another person pass through a doorway, their intent is clear.

Movement communicates awareness. A trustworthy robot should do the same.

It should produce behaviour that shows:

• I noticed you.
• I understand what you may be doing.
• I’ve accounted for your possible movement.
• I’m changing my own action early enough for you to understand mine.

This is more than collision avoidance, it’s behavioural evidence of attention.

Early results from our NVIDIA Cosmos simulations point in this direction. Adding Motion Intelligence signals reduced near-collisions in the most contentious scenario, but the more interesting behavioural changes were that robots yielded earlier, moved more smoothly and reduced planner volatility.

The trust loop

Trust between humans and robots is recursive. Every interaction gives the human new evidence about the machine.

A useful model is:

1. Human signal

A person moves, pauses, hesitates, accelerates, reaches or changes direction.

2. Machine attention

The robot elevates that person within its World State Vector.
A child Motion Intelligence process may begin tracking them more closely.

3. Machine anticipation

The robot estimates likely intent, uncertainty and possible future states.
It considers whether its own planned action intersects with those futures.

4. Machine behaviour

The robot slows, yields, waits, creates space, changes path or communicates.

5. Human interpretation

The person sees evidence that the robot noticed and accounted for them.

6. Trust update

Confidence in the robot rises or falls.

This loop repeats continuously. A single good interaction won’t establish complete trust. But consistent, legible anticipation can build it over time.

The opposite is also true: one interaction in which a robot appears oblivious can undermine confidence quickly, even when no physical harm occurs.

Trust is therefore not a static model output. It’s an accumulated human judgement shaped by the robot’s visible use of attention.

Attention completes the SpatioTemporal architecture

SpatioTemporal Intelligence has previously been framed as a family of capabilities between perception and planning:

Spatial intelligences

• Motion Intelligence
• Intent Analysis
• Human Awareness
• World State Vector

Temporal intelligences

• Future-State Prediction
• Causal Reasoning
• Consequence Modelling
• Planning Handoff

Selective attention connects these capabilities.

Motion Intelligence identifies meaningful change.

Intent Analysis interprets what the change may imply.

Human Awareness establishes social and physical relevance.

The World State Vector maintains the current machine-readable situation.

Attention determines which parts of that state deserve deeper cognition.

Future-State Prediction projects those parts forward.

Consequence Modelling evaluates what may happen under different actions.

Planning acts on that understanding.

System 0 executes the plan.

The complete sequence is:

\text{Perception}
\rightarrow
\text{Motion}
\rightarrow
\text{Intent}
\rightarrow
\text{World State}
\rightarrow
\text{Attention}
\rightarrow
\text{Consequence}
\rightarrow
\text{Planning}
\rightarrow
\text{Action}

From the human side, the sequence continues:

\text{Action}
\rightarrow
\text{Legibility}
\rightarrow
\text{Trust}
\rightarrow
\text{Adoption}

This is why attention belongs inside the technical architecture rather than outside it as a social objective.

Trust is produced through behaviour.

Behaviour is shaped by planning.

Planning depends on consequence.

Consequence requires selective attention.

And attention requires an understanding of motion.

---

Research hypotheses

The theory produces several testable claims.

Motion-informed attention should improve physical safety

Robots that allocate attention using motion and inferred intent should reduce conflict and near-collision rates compared with systems that use proximity or static perception alone.

Selective attention should reduce unnecessary computation

Dynamic allocation of child models should use less inference compute than uniform multi-agent prediction while preserving or improving performance on consequential interactions.

Early attention should produce more legible behaviour

Robots that detect emerging intent earlier should yield sooner, move more smoothly and reduce abrupt replanning.

Legible anticipation should increase human trust

People should report higher trust in robots whose movements visibly account for likely human behaviour before explicit intervention is required.

Human awareness should predict adoption better than task performance alone

For robots operating in shared spaces, perceived awareness of human movement may be a stronger predictor of willingness to purchase or deploy than raw task-completion scores.

Consequence weighting should outperform proximity weighting

In scenarios where the nearest agent isn’t the most consequential, a consequence-weighted system should allocate attention more effectively than distance-based heuristics.

Persistent state should outperform repeated reconstruction

A continuously updated World State Vector should support more stable behaviour than architectures that reconstruct relevance from scratch during each planning cycle.

Dynamic child models should improve ambiguity handling

Creating, deepening and retiring agent-specific inference processes should improve performance in scenes where only a small subset of agents become relevant over time.

These hypotheses provide the foundations for a broader benchmark of trustworthy Physical AI.

That benchmark shouldn’t measure only whether the robot avoids impact.

It should also measure:

• how early it recognises intent
• how quickly it reallocates attention
• whether its attention matches the most consequential agents
• how stable its predictions remain
• how smoothly its behaviour changes
• whether humans can understand its response
• whether people believe the robot noticed them
• whether that belief affects willingness to interact again

---

What this theory doesn’t claim

Motion alone won’t reveal every human intention.

Some situations depend on traffic signals, signs, spoken instructions, gaze, hand gestures, cultural conventions or hidden goals.

A person may move identically towards two objects while intending to interact with only one of them.

A robot may require semantic context to understand why someone is behaving in a particular way.

The theory is not that motion replaces perception, language or planning.

It is that motion provides an essential signal for allocating attention in physical environments.

Nor does selective attention guarantee safety.

Prediction will remain uncertain. Sensors will fail. Humans will behave unexpectedly. Multiple futures will remain plausible.

A trustworthy architecture must expose uncertainty and default towards appropriate caution when its model is weak.

The aim isn’t perfect prediction.

It is earlier recognition of what deserves prediction.

Conclusion

The future of Physical AI won’t be decided only by how many tasks robots can perform.

It will be decided by whether people are willing to let robots perform those tasks around them.

That requires trust.

Trust begins when the machine demonstrates that it has noticed us.

It understands that our movement contains information. It recognises when that movement changes. It increases attention when our possible futures begin to intersect. It considers the consequences of its own actions and responds early enough for us to see that understanding reflected in its behaviour.

This isn’t deep machine cognition.

It’s something more immediate.

A reflex.

An instinct.

A System 1 layer for Physical AI.

The robot doesn’t need to reason deeply about every element of the world. It needs a persistent state, a way to decide what matters, and an architecture that can allocate intelligence accordingly.

Most things can remain in the background.

Some things should be monitored.

A few things deserve prediction.

And the things that could cause harm, conflict or a loss of trust should command the robot’s attention.

Machines don’t need to understand everything equally.

They need to pay attention to what matters.

Thank you for your attention.