The science of Cerebrum
From perception to reasoned action.
A reasoning engine for the physical world.
Neural perception supplies observations with stated confidence. A deterministic symbolic core reasons over them: the same approved knowledge and the same inputs always produce the same decision, and every step can be explained and replayed bit for bit. What is not observed stays unknown, not guessed.
About 15 minutes in full · 3 minutes on the surface · every fold-out is optional
One reasoning architecture. Many physical embodiments.
Live mission model
One mission. A visible chain of reasoning.
01 / Scientific foundations
A connected architecture for robotic cognition.
Perception, knowledge and reasoning have distinct roles. Connected, they turn observations into action.
Neuro-symbolic integration
Seeing acquires meaning through knowledge.
Neural perception reports what it sees with a stated confidence. Domain ontologies relate that evidence to known objects, places and relationships, forming a world view that keeps uncertainty visible. A detected object stays evidence to interpret, not an unquestioned fact.
Knowledge representation
A model of the world. A model of itself.
The world model describes places, objects and relationships. The self-model describes the platform’s size, motion limits, sensors, energy and role. Together they set what a plan can be; model fidelity and freshness bound what the system can conclude.
Model-based reasoning & planning
Memory gives decisions context.
Working memory holds the current situation. Governed Knowledge supplies domain ontologies, policies, models and relevant history. Policies are part of that knowledge, not a check applied afterwards: they decide which candidate actions can be derived at all and how the rest are compared.
Logic-based knowledge synthesis
An update begins as a hypothesis.
Experience suggests explicit, examinable changes to knowledge. Tests, counterexamples and expert review decide whether a proposal suits a defined scope. It becomes knowledge only when experts approve and sign a new version.
Coordination & constrained optimization
Different capabilities. A common objective.
A team of drones shares one world view and divides the work: who maps which sector, on which route, at what time. Assignments are compared on completion time and energy, within each platform’s limits. Coordination alone does not guarantee a globally optimal plan.
Explore the science / the situation lab
One simulation. Five questions.
A team of small drones maps an area in a sandboxed simulation. Pick a question, then change the conditions.
How does an observation become useful knowledge?
A camera frame means little on its own. The domain ontology relates it to objects, places and the task: an obstacle stands in a survey lane; the lane covers a map cell. Reasoning keeps observation, interpretation and conclusion distinct.
Sensor evidence acquires meaning through objects, relationships and context.
What makes an action possible for this drone, here?
A route on the map can still be unsuitable for a particular drone. The world view relates observations to the world model; the self-model states the drone’s sensors, altitude limit and energy. Together they decide which actions are feasible and permitted.
The same area offers different possibilities to different platforms.
How are possible actions compared?
Policies decide which candidates can be derived at all. Each remaining candidate has prerequisites, modeled consequences and an energy cost. Comparing them yields a decision record: what was considered, what was not derivable and why.
Every candidate is examined against the same situation, objectives and policies.
How can experience improve knowledge under expert control?
Experience can reveal a missing rule or an exception. Knowledge synthesis turns it into an explicit proposal that keeps its evidence, is challenged by counterexamples, and enters use only in a new version that experts approve and sign.
Experience can revise a proposal without changing the knowledge in use.
How do several drones contribute to one mission?
The team shares one world view and divides the area into sectors, with deconflicted routes and timing. When conditions change, the plan changes: a drone low on energy hands its sector to a teammate; an unknown cell gets another view.
One world view, divided work, deconflicted routes.
Explore parts of the illustration: tap a numbered point.
Try it · Change the available evidence.
Supported observation
The obstacle has a clear meaning.
A fresh camera frame places an obstacle in the planned lane, which rules the lane out.
- Observed
- An obstacle occupies the survey lane.
- Interpreted
- It breaks the required clearance.
- Next step
- Derive a permitted alternative.
Reasoning trace
- ObservationObstacle in lane
- RelationshipIntersects planned route
- ConstraintClearance not met
- ConclusionExclude this lane
Evidence incomplete
The lane cannot yet be called clear.
Part of the lane is hidden. The world view marks it unknown, not free.
- Observed
- Only part of the lane is visible.
- Unresolved
- Whether the hidden section is blocked.
- Next step
- Request another view of the hidden area.
Reasoning trace
- ObservationPartial visibility
- RelationshipPossible obstruction
- ConstraintClearance unconfirmed
- ConclusionRequest another view
Evidence conflict
Both observations need their context.
An earlier frame shows a clear lane; a newer range reading shows an obstacle.
- Observed
- Two sources disagree about the lane.
- Unresolved
- Whether the scene changed or a source is unreliable.
- Next step
- Reconcile the sources and seek a fresh view.
Reasoning trace
- ObservationSources disagree
- ContextTime and viewpoint differ
- ConstraintNo confirmed clearance
- ConclusionResolve the conflict
Try it · Change the drone or its resources.
Feasible within the model
A high viewpoint is reachable.
The survey drone has the camera, a permitted area and the energy to map and return.
- Capability
- High-altitude mapping.
- Required conditions
- Permitted area and confirmed clearance.
- Feasible task
- Map the cell from the high viewpoint.
Reasoning trace
- EnvironmentHigh viewpoint over buildings
- Self-modelSurvey camera drone
- Resource checkReturn reserve available
- Feasible actionMap from above
Capability changes the plan
The compact drone takes a different task.
Its self-model sets a lower altitude limit: a low-level pass between obstacles, not the high viewpoint.
- Capability
- Low-level mapping.
- Required conditions
- A clear gap and sufficient clearance.
- Outside capability
- The high viewpoint.
Reasoning trace
- EnvironmentLow gap between obstacles
- Self-modelCompact drone, low ceiling
- ConstraintNo high viewpoint
- Feasible actionMap at low level
Resource constraint
A reachable cell may still be infeasible.
The viewpoint is reachable, but the reserve does not cover mapping and the return.
- Capability
- Camera and flight remain available.
- Limiting factor
- Not enough reserve to map and return.
- Next step
- Return; hand the cell to a teammate.
Reasoning trace
- EnvironmentCell remains reachable
- Self-modelReduced energy reserve
- ConstraintReturn requirement unmet
- Feasible actionReturn; hand over the cell
Try it · Change a condition and compare alternatives.
Knowledge shapes the choice
The alternative route meets every condition.
The original lane is blocked; the alternative is permitted, clear and within reserve.
- Relevant knowledge
- Clearance, area and return-reserve policies.
- Original route
- Not derivable: the obstacle breaks the clearance policy.
- Alternative route
- Preferred: permitted, clear and within the energy reserve.
Reasoning trace
- ContextOriginal lane blocked
- KnowledgePolicy + clearance
- ComparisonAlternative meets all conditions
- Preferred candidateAlternative route
Return reserve takes precedence
No survey route fits the energy limit.
Every route needs more energy than the survey-and-return budget allows.
- Relevant knowledge
- A required return reserve bounds every task.
- Original route
- Not derivable: cannot preserve the return reserve.
- Alternative route
- Not derivable: also exceeds the budget.
- Selected response
- Return and hand the cell to a teammate.
Reasoning trace
- ContextEnergy reserve reduced
- KnowledgePreserve required return
- ComparisonSurvey options not derivable
- Selected responseReturn and hand over
Uncertainty limits action
The next useful action gathers evidence.
The original lane is blocked and the alternative’s clearance is not established.
- Relevant knowledge
- Movement requires established clearance.
- Original route
- Not derivable: the obstacle remains.
- Alternative route
- Not yet supported: clearance unknown, even though the route is permitted.
- Request evidence
- Preferred: hold safely and request another view.
Reasoning trace
- ContextAlternative clearance unknown
- KnowledgeEstablish clearance first
- ComparisonMovement not yet supported
- Selected responseRequest another view
Try it · Follow a rule through challenge and revision.
Candidate knowledge
Earlier success suggests a rule.
After several runs, a proposal generalizes: “When the main lane is blocked, use the east lane.” Its scope is too broad.
- Evidence
- Earlier runs used the east lane successfully.
- Assumption
- The lane suits every case where the main one is blocked.
- Status
- Proposal only; the knowledge in use is unchanged.
Reasoning trace
- EvidenceEarlier successful detours
- InductionPropose a general rule
- Open questionWhen does the rule fail?
- StatusCandidate, not approved
Counterexample found
The lane works only under some conditions.
In another run the east lane crosses a closed sector, which policy does not permit.
- New case
- The east lane crosses a closed sector.
- Failure exposed
- Earlier success did not show the lane stays permitted.
- Required response
- Revise or reject the rule.
Reasoning trace
- EvidenceSector closed
- TestApply the proposed rule
- ConflictRoute not permitted by policy
- StatusReturn for revision
Refined proposal
Expert judgment makes the limits explicit.
Use the east lane only when policy permits it, clearance is confirmed and the reserve holds. Otherwise hold or return.
- Safeguards
- Permitted area, clearance and return reserve.
- Next step
- Validate the revision and record the expert decision.
- Status
- A proposal until experts approve and sign a new version.
Reasoning trace
- Expert inputDefine conditions and fallback
- RevisionNarrow the rule
- ValidationTest the revised scope
- StatusAwait approval and signing
Try it · Change the team’s situation.
Divided work
Each drone maps its own sector.
One drone maps the west sector, a teammate the east, on deconflicted routes.
- West drone
- Maps the west sector.
- East drone
- Maps the east sector.
- Shared plan
- Deconflicted routes and timing; one shared map.
Reasoning trace
- Task AWest sector → west drone
- Task BEast sector → east drone
- CoordinationShare progress and timing
- CompletionMerge both maps
Task handover
One drone returns; a teammate takes over.
The east drone returns to the home pad and hands its remaining cells to the west drone.
- East drone
- Returns to the home pad within its reserve.
- West drone
- Takes over the remaining east cells.
- Shared plan
- Updated assignments; no cell left uncovered.
Reasoning trace
- TriggerEast reserve below plan
- AllocationRemaining cells → west drone
- CoordinationDeconflict the new route
- CompletionWhole area still covered
Unknown stays unknown
An unobserved cell triggers another view.
One cell was never observed. It stays unknown, not assumed clear, until the east drone views it.
- East drone
- Views the unknown cell, then continues its sweep.
- West drone
- Continues its sector unchanged.
- Shared update
- Revised order and expected coverage.
Reasoning trace
- New objectiveObserve the unknown cell
- AllocationNearest suitable drone
- CoordinationUpdate order and timing
- CompletionTrack the revised sequence
Go deeper: evidence, symbols and uncertainty
- Neural perception
- Learned models extract patterns such as objects or features from observations. Their outputs still need context and a stated confidence.
- Symbolic representation
- Entities, attributes and relationships express what the evidence refers to, so rules and models can operate on it.
- Evidence and belief
- A sensor reading is evidence. An interpretation is a belief supported by that evidence, which may change as further observations arrive.
A graph makes relationships explicit; it does not, by itself, establish that every represented statement is correct.
Scientific foundation
Neurosymbolic AI: The 3rd Wave ↗
Garcez & Lamb · Neural learning connected with symbolic knowledge and reasoning. Scientific context, not independent validation of Cerebrum.
Recognizing an object and understanding its significance are connected but distinct.
Go deeper: state, capability and feasibility
- World model
- Geometry, places and relationships describe the operating world. A map is one part of this representation.
- Current world view
- Recent observations describe what appears to be happening now, including changes and unresolved uncertainty.
- Operational self-awareness
- The platform represents its own state, capabilities and limits, and these shape which actions it can consider.
Model completeness, accuracy and freshness limit the conclusions that can be drawn. An action feasible in a model still depends on the validity of that model.
Scientific foundation
Models: Introduction to Robotics and Perception ↗
Dellaert & Hutchinson · Representing a robot and its environment. Scientific context, not independent validation of Cerebrum.
Feasibility is a relationship between the task, the environment and the specific platform.
Go deeper: planning, memory and modeled consequences
- Prerequisites and effects
- An action model states what must hold before an action and how the represented state may change afterward.
- Memory with different roles
- Current context, Governed Knowledge and relevant history contribute different information to the same reasoning process.
- Bounded comparison
- The candidates, prediction horizon, objective and policies define the scope of the comparison. A choice is assessed within that scope.
A selected action is not proof of a globally optimal outcome. Repeatable computation and a faithful model are separate properties.
Scientific foundation
Planning Algorithms ↗
Steven M. LaValle · States, actions, feasibility, planning and constraints. Scientific context, not independent validation of Cerebrum.
A useful explanation reveals the alternatives, assumptions and constraints behind a decision.
Go deeper: deduction, induction and abduction
- Deduction
- Given a policy that keeps drones out of closed sectors and a lane known to cross one, conclude that the lane is not permitted.
- Induction
- Repeated successful detours may suggest a more general rule. Further cases can expose where that generalization fails.
- Abduction
- A new obstacle may suggest an explanation, such as a temporary structure in the simulation. It remains a hypothesis until evidence supports it.
Deterministic synthesis makes the same proposal reproducible under fixed inputs and configuration. It does not make a generalization universally true or approve its use.
Scientific foundation
Inductive Logic Programming at 30: A New Introduction ↗
Cropper & Dumančić · Learning explicit logical hypotheses from examples. Scientific context, not independent validation of Cerebrum.
Experience suggests an improvement. Experts decide whether it becomes knowledge.
Go deeper: allocation, dependencies and coordination
- Task allocation
- Match tasks to eligible drones under capability, availability and energy constraints.
- Dependencies and timing
- A task may need another result first. Sequencing and deconfliction are part of the shared plan.
- Communication and uncertainty
- Assignments depend on what the team knows. Missing status or delayed observations call for reassessment, not assumed success.
Claims of improved efficiency need a defined baseline, common mission conditions and measured outcomes. This illustrative assignment is not a performance benchmark.
Scientific foundation
A Formal Analysis and Taxonomy of Task Allocation in Multi-Robot Systems ↗
Gerkey & Matarić · Formal analysis of multi-robot task allocation. Scientific context, not independent validation of Cerebrum.
Shared objectives become coordinated action through capabilities, information, timing and energy.
Illustrative science model of a sandboxed simulation. It explains relationships and constraints; it does not run a Cerebrum system or report measured performance.
02 / Governed Knowledge synthesis
Experience informs. Experts govern.
Cerebrum reasons with Governed Knowledge: approved, versioned and signed.
People govern the knowledge, not every decision. Deterministic synthesis turns experience into proposed updates linked to their evidence. Experts challenge them, then approve and sign a new version.
Experience → proposed knowledge → expert review → signed version
What Governed Knowledge contains
- Domain ontologiesWhat exists in the operating world and how it relates: objects, places, capabilities, relations.
- PoliciesWhat is permitted, required or forbidden, limits of authority, and when a person must confirm. A forbidden action is simply not derivable.
- Operational knowledgePlatform models, procedures and relevant history.
Where a policy requires human confirmation, the engine asks for it, as a policy outcome written in the knowledge. Execution stays with each platform’s own controllers and safety systems.
From an observation to a signed update.
Five stages of the knowledge cycle. Press run to send the example lesson through them.
01 / Operational evidence
Keep the experience in context.
Record observations, decisions and outcomes with the knowledge version and conditions. Keep exceptions, so one success is not mistaken for a rule.
What moves forward
A traceable evidence record.
Evidence only · no knowledge change
02 / Logical synthesis
Make the proposed change explicit.
Relate the evidence to existing ontologies, policies and models. Derive what follows; form hypotheses where a rule seems missing.
What moves forward
A candidate update with its evidence and reasoning.
Candidate knowledge · not in use
03 / Tests & counterexamples
Ask where the proposal breaks.
Replay cases under varied conditions. Look for conflicts, weak assumptions and unacceptable decisions. Failures send the proposal back.
What moves forward
A validation record: cases examined, failures found, limits still open.
Under evaluation · passing tests does not activate knowledge
04 / Human expert review
Turn domain expertise into safeguards.
Experts approve, revise or reject, and set scope, limits and fallback. If an action class needs a person to confirm it, they write that into the policy.
What moves forward
An expert decision with explicit scope and reasons.
Approval is a review decision · commissioning is a separate step
05 / Controlled commissioning
Put a signed version into use.
Release the signed version for the intended platform, task and scope, with its history attached.
What moves forward
A versioned, signed release linked to its evidence, tests and approval.
Commissioned knowledge · bounded by its approved scope
Illustrative example / Simulated aerial survey
A useful lesson. With explicit limits.
In the simulation, a drone keeps finding its survey lane blocked. An alternative lane has worked in earlier runs. What should the system learn?
Initial proposal
“When the survey lane is blocked, use the east lane.”
Earlier success supports a hypothesis. It does not show the lane is permitted or suitable in every case.
↓ Tests and expert review: revise the rule and its scope
In one run the east lane crosses a closed sector. The validation record keeps that counterexample.
Refined proposal with safeguards
“Consider the east lane only when policy permits it, current observations confirm clearance, and the drone keeps its required energy reserve.”
Otherwise the drone holds or returns. The revision is validated before approval.
Commissioned
Knowledge version 2, signed and linked to its evidence, tests and approval.
Use stays bounded by the approved scope
↳ New experience starts the next cycle. Every proposal returns to review.
The thread that makes it auditable
Evidence, reasoning, tests, expert decisions and the signed version stay linked, so any update can be explained.
What does “deterministic” mean here?
Given the same approved knowledge and the same inputs, the engine always reaches the same decision, and synthesis always produces the same proposal. Both can be replayed bit for bit. This does not make incomplete evidence certain or a generalization universally true; testing and expert judgment remain essential.
03 / Evidence & research
Claims should be open to examination.
Foundations explain the approach. Reproducible experiments show what an implementation can do, and under which conditions.
These are proposed validation questions. No Cerebrum benchmark results are presented in this draft.
Foundational reading
Research that helps explain the concepts behind the architecture.
Neural + symbolic reasoningArtur d’Avila Garcez & Luís C. Lamb · 2020Neurosymbolic AI: The 3rd Wave
Integrating neural learning with symbolic knowledge representation and logical reasoning. Read paper →
Learning logical hypothesesAndrew Cropper & Sebastijan Dumančić · 2022 revisionInductive Logic Programming at 30
An introduction to learning logical rules from examples, including limitations and open questions. Read paper →
Explicit world models & planningRichard E. Fikes & Nils J. Nilsson · 1971STRIPS: A New Approach to Problem Solving
A foundational account of planning through operators that transform a represented world state. Read paper →
These publications provide scientific context; they do not independently validate Cerebrum or imply endorsement by their authors.
04 / Ecosystem
Built with partners. Grounded in operations.
Cerebrum is the shared reasoning and knowledge layer. OEMs bring platforms, integrators build solutions, end clients define the work.
Cerebrum technology
The shared reasoning & knowledge layer
- World & capability models
- Governed Knowledge: domain ontologies and policies
- Replayable decision records
Choose a corner of the triangle to read that role.
The robotic platform
OEMs
Original equipment manufacturers
Bring the machines and the engineering behind physical action.
- Provide robots, sensors, actuators and control systems.
- Define capabilities, interfaces and equipment limits.
- Keep execution in the platform’s own controllers and safety systems.
With Cerebrum Machine capabilities and limits become explicit knowledge the engine reasons with.
The operational solution
Systems integrators
Connect technology with the use case
Turn platforms and workflows into one working solution.
- Connect robotic and enterprise systems.
- Configure domain ontologies, policies and mission workflows.
- Validate the integrated solution and support commissioning.
With Cerebrum A common reasoning layer gives different machines a shared operational context.
The operating mission
End clients
Organizations that deploy and operate
Bring objectives, domain expertise and accountability.
- Define outcomes, operating policies and acceptance criteria.
- Provide experts who review, approve and sign knowledge versions, including policies.
- Authorize deployment and oversee operational performance.
With Cerebrum Operational intent becomes Governed Knowledge, and every decision based on it can be examined.
Together in practice / Conceptual example
An industrial inspection mission.
A conceptual illustration, not a deployed system. An OEM supplies the inspection robot; an integrator connects it to the facility’s ontology; the client’s experts approve inspection priorities and access policies as part of the Governed Knowledge. Cerebrum reasons with that context and keeps a replayable record of each decision.
↳ Experience returns as proposed knowledge. Partners and domain experts review it, then approve and sign a new version before use.