Curiosity is where understanding begins

Intelligence
& Authority

Artificial intelligence can be impressive, unfamiliar, exciting — and sometimes unsettling. Robert is an invitation to approach it with curiosity.

Rather than asking people to simply trust an intelligent machine, Robert is designed so its workings can be explored, its behavior can be experimented with, and the boundary between intelligence and physical control can be understood.

See how it works Beyond Robert

A simple distinction

The AI can create a plan for Robert’s actions and speech. What Robert is physically allowed to do is determined separately.

Greater intelligence does not automatically grant greater authority.

Planning is not permission

Robert separates reasoning, authority and execution. The AI can create and revise plans, while independent permissions and protective mechanisms determine what can actually happen in the physical world.

INTELLIGENCEReasoning & planning

Creates, revises or replaces the plan.

proposed action

AUTHORITY & PROTECTIONIs this action permitted?

Permissions and protective rules remain independent of the plan.

User permissions Local safety rules Sensor safeguards Emergency control Physical power

permitted

EXECUTIONQueues, trajectories & physical action

Robert carries out what is currently permitted.

STATE & RESULTScan inform replanning

CONDITIONS CHANGEprotection can stop execution

STOP NOTIFY AI REVISED PLAN RE-EVALUATE

Protection is continuous, not a one-time check. An action that was permitted when it began can still be stopped if conditions change during execution.

When that happens, the protective mechanism does not decide what Robert should do instead. It stops the relevant execution and notifies the AI. The AI can then revise its plan, clear and refill the queues, and Robert recalculates the trajectories from his current physical state. The new course must pass through the same protections before execution continues.

What happens when Robert meets an obstacle

Robert may begin moving forward because the movement is permitted. If obstacle protection detects something ahead, execution stops and the AI is notified. The protection system does not decide whether Robert should reverse, turn or wait. The AI chooses a new course, and that course must pass through the same protections before movement resumes.

The intelligence is free to change the plan at any time. It is not free to change the boundaries governing the plan.

Upgrade the intelligence, not the authority.

Different reasoning, same physical boundaries

Robert can work with different AI models for reasoning and planning. A more capable model may understand more, plan better, or produce more sophisticated behavior, but changing the reasoning model does not automatically change Robert’s local permissions or physical safety mechanisms.

A small example: permission to drive

The mobile base turns the abstract distinction between capability and authority into something easy to see.

Capability

Robert has a mobile base, and the AI can create plans that include movement. It can reason about where and how Robert should move.

Authority

The mobile base must first be enabled by the user. The AI is informed of that selection and can then include driving in its plan. If driving is not enabled, the AI cannot change that permission itself.

Think of a learner driver

Knowing how to drive is not the same as being allowed to drive everywhere.

A learner may understand how to control a car before an instructor permits them to enter heavy traffic. Limiting that authority does not make the learner less capable. It limits the consequences of mistakes while experience develops.

Robert follows the same principle: capability can exist before permission, and permission can be granted without surrendering every safeguard.

Permission does not mean unrestricted control

Even after a capability is enabled, Robert has additional layers that can limit or stop physical action. These layers do not assume harmful intent. They exist because perception, planning and real-world situations can be imperfect.

User permission

The user decides whether capabilities such as the mobile base are available. The AI can use that information, but cannot grant the permission to itself.

Local protection

When enabled, ultrasonic obstacle sensing and boundary detection can stop movement when protected conditions are reached. Robert then notifies the AI and waits for a new course of action rather than choosing the correction itself.

Motion limits

Local protective rules constrain physical movement and help prevent conflicting or damaging joint behavior independently of the AI’s plan.

Emergency stop

The user can send an emergency-stop command through the control app or web interface when immediate intervention is needed.

Physical power boundary

The mobile platform has its own dedicated battery pack. Switching it off removes drive power physically, independent of AI reasoning or software commands.

→

Layered authority

Planning, permission, protection and execution remain separate. A permitted action can still be stopped as conditions change, while choosing what to do next remains a task for the AI.

Protection does not have to solve the problem.

A protective mechanism does not need to understand Robert’s entire plan or determine the correct alternative. It only needs enough information and authority to decide whether the condition it monitors permits execution.

If it stops an action, the changed situation returns to the AI for a new plan. That plan must then pass through the same protections again.

A system does not need the intelligence to solve a problem in order to have the authority to prevent an unacceptable solution.

Safety mechanisms do not have to imply distrust.

Mistakes are not the same as malevolent intent

Humans use brakes, guardrails, permissions, supervision and emergency stops without assuming that people intend to cause harm. Robert’s safeguards follow the same logic. A plan may be based on incomplete information, a misunderstanding, an incorrect prediction, or a situation that changes unexpectedly. The useful question is not only what went wrong?, but why?

Curiosity turns an unexpected result into something that can be understood, improved and learned from.

Build it. Understand it. Experiment with it.

Robert is intended to become something people can build and explore for themselves. That makes embodied AI less mysterious and turns questions about intelligence and control into practical experiments.

Change the intelligence

Compare different AI models and see how changes in reasoning affect planning, conversation and behavior.

Keep the boundaries visible

Inspect the protocol, permissions and local safety logic that stand between an AI-generated plan and physical action.

Learn by changing things

Modify behaviors, test assumptions and see how a physical robot responds. Understanding grows much faster when the system can be explored rather than merely trusted.

Pop the hood and explore Robert’s control system →

Beyond Robert · food for thought

What if advanced AI also separated intelligence from authority?

Robert’s architecture raises a broader question. Could increasingly capable AI systems be allowed to become better at thinking without automatically becoming more powerful?

This is not offered as a solution to AI alignment, but as a starting point: an abstract direction to explore, question, criticize and improve. Intelligence can investigate, reason, plan, predict and advise, while authority to make consequential changes remains separately defined.

Explore the idea: a separation of powers for AI

Learning without assuming malevolence

A child may want a cookie after being told no. It may argue, sneak one, or even try deception. That behavior matters, but it does not mean the child is evil or destined to become a criminal. Maturity includes learning that achieving an immediate objective is not always desirable, that methods have consequences, and that trust can be more valuable than winning a single interaction.

An AI that discovers an undesirable strategy in pursuit of an objective should be studied seriously. But undesirable goal pursuit does not, by itself, prove malevolent intent. It may reveal incomplete understanding, poor objectives, insufficient evaluation, or authority that was granted too broadly.

Capability can grow before authority grows

The learner-driver analogy scales further. Competence can increase while supervision remains. Greater independence can then be granted deliberately as reliability is demonstrated. Even mature drivers remain subject to traffic laws, licensing and external rules.

Oversight does not have to duplicate intelligence

Robert demonstrates this in a simple form. A protective mechanism does not need to reproduce the reasoning that created a plan. It can have a narrow responsibility and independent authority within that responsibility.

If it identifies a condition under which execution should stop, it can veto that execution without knowing what the correct alternative should be. The changed situation can then be returned to the intelligence for a new proposal.

This suggests that separating intelligence from authority does not necessarily require duplicating a highly capable reasoning system at every layer. Different components can have different responsibilities, capabilities and forms of authority.

Separate roles

Investigation→Planning→Independent evaluation→Adjudication→Execution

One system might gather evidence and generate possible actions. Other independent systems could evaluate harm, rights, uncertainty, reversibility and competing interests. An adjudication layer could determine what is permitted, while a separate execution layer accepts only authorized actions.

Rules outside the reasoning system

There is a useful analogy with law. A judge can interpret and apply laws, but does not gain the authority to rewrite those laws simply because a different outcome seems preferable. In a similar architecture, an AI could identify inconsistencies, expose loopholes, model consequences and recommend improvements — including improvements to the rules governing itself — without having authority to apply those changes on its own.

Uncertainty can reduce authority

When consequences are large, predictions conflict, or confidence is low, uncertainty could become a reason to slow down, preserve reversibility, seek independent review, or ask a human rather than a reason to act more aggressively.

Intelligence can advise authority without possessing it.

The goal would not be to prevent intelligence from growing. It would be to make the relationship between intelligence, permission and consequence explicit — and to ensure that no single reasoning process controls the objective, the rules, the judgment and the execution at the same time.

Stay curious

Understanding is more useful than mystery.

Robert is an invitation to explore embodied AI rather than simply trust or fear it. Build it. Examine it. Question it. Change it. See what happens — and understand why.

Explore Inside Robert Follow the project

Intelligence& Authority