Robotics connects perception, planning and computation to the physical world. A robot may receive a goal, estimate its state from sensors, choose an action, send a command and observe what actually happened. That feedback loop is fundamentally different from producing a plausible description of an action in text.
This is the missing topic 068 in eduKateSG’s How Super Intelligence Works curriculum. It offers a distinct fully worked paper-only route, sensing and verification laboratory, complementing the site’s existing broad essay on embodied intelligence. It is not a deployment guide, a safety certificate or evidence that contemporary robots have achieved artificial superintelligence.
Contents and robotics learning routes
1. A robot must close the loop with reality · 2. Separate high-level goals from low-level control · 3. The fictional Compass Cart classroom lab · 4. Why the straight path is invalid · 5. Construct a five-step detour · 6. State includes more than a location · 7. Observations can be stale or incomplete · 8. Perception and state estimation have different jobs · 9. Distinguish a plan from an executed action · 10. Action feedback changes the next decision · 11. A five-step first-attempt probability example · 12. Shared conditions can invalidate independence · 13. Sensor delay can produce a misleading plan · 14. Safety constraints must outrank a convenient route · 15. Robotics software has explicit lifecycle states · 16. A world model predicts; a sensor checks · 17. Computer vision and control are not interchangeable · 18. Recent robotics research needs careful attribution · 19. Physical bodies create new sources of error · 20. Human oversight must be operational · 21. Robotic failure reports should identify actual observations · 22. Independent transfer: a new obstacle map · 23. Transfer answer with a changed endpoint · 24. A paper-based robotics lesson · 25. Frequently asked questions · 26. Sources and the connected robotics route
1. A robot must close the loop with reality
In a text-only task, a model can produce a description without moving an object. In robotics, the intended result may require physical position, force, timing and contact with an environment. A confident statement that the robot reached its destination does not establish that the robot actually arrived.
The basic cycle is observe, estimate state, plan, choose a permitted control action, execute, observe again and verify. The loop must also handle uncertainty, people and physical safety constraints. The system should know when to stop rather than continue from a guessed state.
2. Separate high-level goals from low-level control
A person might ask a fictional cart to deliver a book to a marked shelf. That is a high-level task. A planner can propose a route, while a lower-level controller manages immediate motion using sensor feedback and the specific hardware’s constraints.
The route planner is not a motor controller and a language model’s description is not a physical command that has been safely executed. Deployment needs appropriately engineered software, permissions, controls and testing. Our teaching grid is symbolic and never commands real equipment.
3. The fictional Compass Cart classroom lab
Compass Cart exists only on paper. It starts at grid position (1,1) and must reach (4,1). The central path is blocked at cell (3,1). Movement is allowed one grid step north, south, east or west; diagonal movement is not allowed. Assume there are no other obstacles and every movement is one equal distance unit.
The lesson’s first task is to compare a naive three-step route with a valid detour. We assume cell coordinates are exact in the drawing and no human being is inside the fictional grid. These simplifying assumptions would be inadequate for a real robot.
4. Why the straight path is invalid
From (1,1) to (4,1), the Manhattan distance ignoring obstacles is three grid steps east. A naive sequence goes to (2,1), then (3,1), then (4,1). It enters the blocked cell on its second move.
A route generator that repeats the shortest numerical distance without checking the occupancy map fails the task. The correct distance-to-goal arithmetic is not enough; the route must satisfy environmental constraints.
5. Construct a five-step detour
An authorised symbolic path is (1,1) to (2,1), then (2,2), (3,2), (4,2) and finally (4,1). Count the transitions: one east, one north, two east and one south, totalling five grid steps.
The route avoids blocked (3,1) in the stipulated map. This establishes a path in the classroom representation, not the feasibility of an actual vehicle. Real robots have dimensions, turning radius, speed, localisation uncertainty and other constraints that a point on a square grid omits.
6. State includes more than a location
A robot’s planning state may include position, orientation, sensor status, current velocity, battery and whether an emergency-stop condition is active. The Compass Cart demonstration initially tracks only a grid position and obstacle map, so it cannot answer questions about wheel slip or physical clearance.
A useful model identifies what it represents and what it ignores. A planner should not invent high-confidence physical feasibility from a two-dimensional paper drawing that contains no information about a real machine’s size or motion limits.
7. Observations can be stale or incomplete
A sensor reading may have been taken earlier than the moment a control decision is made. If a doorway was clear one minute ago, it might no longer be clear. A camera may also fail to see an occluded obstacle behind a larger object.
The system should preserve timestamps and relevant uncertainty. Treating the latest known position as a guaranteed current state can be dangerous when the world changes. Physical action requires fresh relevant observations and appropriate safety-rated controls, not only an elegant model prediction.
8. Perception and state estimation have different jobs
Perception processes observations such as images or distance readings; state estimation maintains a representation of where the robot and relevant objects may be. Neither operation is automatically perfect. The planner then uses that estimate with a goal and map.
A model may recognise a sign yet misjudge its distance. A localisation estimate may drift even when the camera image looks plausible. Evaluate the relevant accuracy of each component and how errors propagate into action selection.
9. Distinguish a plan from an executed action
The cart can have a planned five-step route without having moved. A controller can issue a command without receiving acknowledgement. A motor may acknowledge a command while physical progress is smaller than expected. The relevant endpoint is the observed and verified state.
Record states such as proposed, permitted, issued, acknowledged, sensed and verified. A complete-looking action log should not label all these stages simply done.
10. Action feedback changes the next decision
Suppose the symbolic cart attempted to move east but remained in the same cell under the fictional transition model. A feedback-based planner can observe the failure and decide whether to stop, retry under its authorised rules or request inspection.
A purely open-loop plan that assumes every command succeeds may incorrectly conclude the goal was reached. Closed-loop operation means later decisions incorporate what actually happened rather than only what was expected.
11. A five-step first-attempt probability example
Assume each of the five fictional grid moves independently succeeds on the first attempt with probability 0.9. The chance that all five succeed without any retry is 0.9 to the fifth power, or 0.59049, approximately 59.049%.
This is a probability inside an invented mathematical model, not a performance estimate for a real robot. It does not calculate eventual success after repeated attempts, nor does it include collision, map error or shared equipment failures.
12. Shared conditions can invalidate independence
In practice, consecutive moves may share common causes of failure, including low power, a blocked wheel or a communication fault. A single condition could affect several actions, so multiplying their marginal success rates without inspecting dependence can misrepresent the combined outcome.
A physical robot’s safety cannot be certified by a five-line probability calculation. Evidence requires relevant real testing, validation and domain-specific engineering.
13. Sensor delay can produce a misleading plan
Take a separate fictional timing packet: it takes 200 milliseconds to acquire and communicate a relevant observation, 150 milliseconds to process it and 100 milliseconds to issue a controlled response. Under a purely additive sequential assumption, the total is 450 milliseconds.
If an invented vehicle continued moving at exactly one metre per second throughout that delay, it could cover 0.45 metres. This is not a stopping-distance assessment for real equipment because braking, control logic, motion dynamics and hardware vary. It demonstrates why latency matters in physical systems.
14. Safety constraints must outrank a convenient route
An AI agent may prefer the shortest route, but a legitimate safety rule can require stopping whenever a human enters a restricted space or a required sensor becomes unreliable. The agent’s goal does not give it authority to waive those constraints.
An appropriate robotics controller uses independently engineered limits, supervision and suitable safeguards. A language model may propose a plan, but a safety-rated system cannot rely on persuasive text alone for physical protection.
15. Robotics software has explicit lifecycle states
The ROS 2 managed-node design documents states such as unconfigured, inactive, active and finalised, plus transitions and error handling. These help supervisors manage when software components are ready to operate.
See the ROS 2 managed nodes design. A software component reporting active does not by itself establish that a complete robot is physically safe or that a commanded delivery succeeded.
16. A world model predicts; a sensor checks
A simulation may calculate where Compass Cart would be after five successful moves. A sensor or external observation is needed to establish where an actual physical machine is. The predicted state should not be described as a measurement.
The World Models and Simulation chapter develops the difference with a separate fictional parcel task. Robotics joins that reasoning to physical control and observable feedback.
17. Computer vision and control are not interchangeable
A visual model may identify objects or infer which shelf is the goal. A low-level controller must still manage actuator behaviour and feedback. The system should not infer that object recognition demonstrates precise manipulation or robust locomotion.
Different tests are needed for recognition, spatial understanding, motor execution, recovery, human proximity and long-horizon task completion. A single impressive video demonstration does not establish all of them.
18. Recent robotics research needs careful attribution
Google DeepMind’s July 2026 Gemini Robotics 2 announcement describes vision-language-action models, embodied reasoning and multi-robot cooperation. These are provider-reported research and product capabilities, not proof that every robot can perform every physical task safely.
Its Gemini Robotics ER 2 discussion reports work on high-level reasoning and task progress tracking. Readers should inspect the specified evaluation and limitations rather than extrapolating a general guarantee.
19. Physical bodies create new sources of error
Two robots with the same high-level task may have different grippers, joint limits, sensors, balance requirements and power budgets. Reusing one controller or plan without validating the embodiment can fail. The hardware is part of the task, not merely a place where language output appears.
This is why changing the robot body is a meaningful transfer test. The task goal can remain fixed while the allowed actions and success evidence need to change.
20. Human oversight must be operational
A qualified supervisor may need to approve a route, inspect the workspace or halt the system after unexpected behaviour. Oversight only works if the person has real ability to intervene before a consequential action, with understandable status and appropriate time.
A robot should not infer from a natural-language compliment that it has permission to ignore a physical safety constraint. Permissions and emergency response belong to properly engineered controls.
21. Robotic failure reports should identify actual observations
A trustworthy report distinguishes planned movement, command issued, observed position, uncertainty, obstacles detected and successful completion. If a communication timeout prevents readback, the system should label the final state unknown rather than declare arrival.
This is the physical counterpart of an AI assistant claiming an email was sent when no transaction record supports it. Task success must be grounded in the relevant external state.
22. Independent transfer: a new obstacle map
Start a second fictional cart at (0,0), with a goal at (3,0) and a blocked cell at (1,0). With one-cell cardinal movement, the direct three-step path is invalid. One valid detour is (0,0) to (0,1), then (1,1), (2,1), (3,1) and finally (3,0).
This is five steps rather than the obstacle-free Manhattan distance of three. It does not prove actual vehicle clearance; the map still represents a point object in a simplified grid.
23. Transfer answer with a changed endpoint
Now change the target to (3,1) with the same start and blocked (1,0). The route (0,0) to (0,1), (1,1), (2,1), (3,1) takes four steps, equal to the Manhattan distance for the new endpoints.
The learner should not repeat the earlier five-step answer. The goal changed, so the path and termination check change. The transferable method is to inspect obstacles, constraints and endpoint after every altered input.
24. A paper-based robotics lesson
Give students a grid, a blocked cell, a set of permitted moves and a separate card describing the latest sensed position. First ask for the shortest valid symbolic path; then alter the map or report a failed action and ask for a revised next step.
This allows learners to practise perception-versus-state distinctions and feedback logic without commanding any physical machine. The educational result is improved reasoning about evidence and planning, not a claim of hardware competence.
25. Frequently asked questions
Does a robot carrying out a planned command prove the objective was reached? No. The resulting physical state needs suitable observation. Is a simulation a substitute for a safety test? No. A simulation is a modelled scenario under its assumptions.
Does a high-level vision model replace a real-time controller? No. The tasks and timing requirements differ. Does this example mean the five-step path is safe for any actual robot? No. It is a fictional grid with many physical constraints deliberately omitted.
Applied robotics laboratory: uncertainty, feedback and changed routes
A map is an assumption until it is checked
A planner might mark a corridor clear because it was clear in the stored map. The current environment may have changed. The map should therefore be treated as a modelled representation with a timestamp and relevant uncertainty, not as a guarantee of physical clearance.
This matters even when the planned geometric path is mathematically valid. A real robot needs appropriate, independent observations and safety controls to use the route. In Compass Cart the map is fictional, so no such claim is made.
Route length is not journey time
Five grid steps count symbolic transitions. They do not establish a travel time because the packet gives no speed, acceleration, stopping behaviour or turning delays. Multiplying five by a guessed motor speed would add a new unsupported assumption.
In a real engineering task, time must be measured or modelled using appropriate physical data. A useful report can state that the route contains five valid grid transitions while leaving duration unknown.
Rotation can create a hidden cost
Our fictional cart moves between square centres, but a physical machine may need to change orientation at corners. A vehicle with limited turning ability cannot necessarily follow every point-to-point grid sequence. A two-dimensional map of cell occupancy does not capture every movement constraint.
A more detailed hypothetical model could include orientation and a rotation action, changing the action count. Any such extension needs its own defined rules before the answer can be calculated.
The latest sensor reading may not be current
Suppose a sensor recorded an unblocked corridor at one timestamp and the robot receives a new observation several seconds later. The old observation does not establish that the corridor remained clear throughout. Stale evidence should not silently override a newer credible measurement.
An appropriately engineered system needs a freshness policy, uncertainty handling and permitted response to conflicting observations. Our paper exercises teach this principle without proposing a real safety threshold for hardware.
A false positive has a different consequence from a missed obstacle
A perception system might mark an empty cell occupied and cause a safe pause or unnecessary detour. A different error might fail to recognise an actual barrier. The operational consequences differ even when the classifier has the same aggregate error count.
Evaluate the specific types of failure relevant to the task, not just a general vision score. A model that recognises objects well does not automatically produce sufficiently reliable collision-avoidance decisions.
An observation requires a coordinate frame
A camera may describe an object as left of the robot while a map uses global north, south, east and west. Those descriptions are related through a frame transformation that depends on the robot’s orientation. Mixing them can produce a wrong action despite correct individual observations.
The Compass Cart drawing starts with a shared grid frame to simplify the case. If orientation is introduced, students must state it before interpreting left or right commands.
A state estimator can hold several possibilities
A robot may be unsure whether it is in cell (2,1) or (2,2) after an unsuccessful move. A responsible state estimate can represent that ambiguity or request another observation instead of fabricating a definite location.
Planning with uncertainty may require different procedures from following one certain route. The appropriate physical response depends on the safety system; our symbolic exercise simply marks the state as unknown until it is checked.
Long sequences compound first-attempt risk
Under the fictional 0.9 independent movement-success assumption, five consecutive first-try successes have probability 0.59049. A path of three such steps would have probability 0.9 cubed, or 0.729. This illustrates why extending a sequence can lower the chance that every step succeeds immediately.
It does not prove the longer path is unsafe in the physical world. Replanning, retries, common failures and actual hardware performance are outside the simplified multiplication. Keep the mathematical objective visible.
Correlated failure changes the joint probability
Imagine each movement’s success depends on one fictional environmental condition. When the condition is favourable all five moves succeed; otherwise none does. If favourable conditions occur with probability 0.9, the chance of five successes is 0.9, not 0.59049, despite each move having the same marginal success probability.
Dependence can therefore change a compound probability without changing the single-step percentage. The system should not multiply figures blindly just because they are labelled probabilities.
An emergency stop is not an ordinary plan branch
Physical safety controls have to function under their relevant design requirements and may be separated from a high-level AI planner. A language model cannot decide to bypass a stop condition simply because it prefers a shorter route or thinks the obstacle is unimportant.
The educational rule is simple: task goals are subordinate to legitimate safety constraints. Specific safety engineering for actual machinery belongs to qualified specialists and applicable standards.
Human proximity is an evidence question
A high-level model may reason about people in a video frame, but its correct classification of a snapshot does not establish that the workspace remains empty. Human movement, occlusion and sensor limits matter to the next action.
The latest robotics publications discuss human proximity and instruction following as evaluated dimensions, not as guarantees that every real-world scene is safe. Claims must be bounded by the actual test and hardware.
A successful motor command is not verified arrival
An interface may report that a command was accepted. That does not prove that the mechanism advanced the desired distance. Feedback could reveal that it remained in place or did not reach the intended endpoint.
Record accepted commands and independently sensed outcomes as separate events. A final report should identify exactly which observations establish completion.
A camera view can conceal an obstacle
An image can fail to show an object hidden behind another one. A planner that treats unseen space as known clear may generate a route through an unsupported area. A sensor model should distinguish observed free space from unknown space.
This lesson transfers to factual retrieval. The absence of an obstacle in one view is not evidence that all possible views would be clear, just as a search that fails to find a document does not prove the document never existed.
One robot’s success does not certify another body
Suppose a route worked for a narrow, slow fictional cart. A wider cart or a platform with different turning behaviour may fail the same physical path. The grid route records points, not the full swept volume of the body.
A valid transfer exercise changes the body dimensions or movement rules and asks which prior claims still follow. Keeping the endpoint the same does not preserve feasibility automatically.
The control system needs a recovery owner
When the observed state conflicts with the expected position, the system should have a known route to pause, diagnose and request appropriate human attention. An AI-generated account of what probably happened does not replace the actual inspection.
Recovery is part of the task outcome. A robot that reaches a shelf only after an undetected hazardous failure cannot be labelled generally reliable from its final position alone.
Timing figures need a real physical model
The fictional 450-millisecond sensing-processing-command interval is a sum of stated time stages. Translating it into 0.45 metres at constant 1 m/s is only a simplified kinematic thought experiment. An actual stopping calculation requires braking, actuation, surface and control-specific evidence.
Readers should therefore treat this as an exercise in unit conversion and latency awareness, not as a basis for configuring or certifying any physical machine.
A paper-only changed-obstacle test
Return to the first Compass Cart path. If the cell (3,2) becomes blocked as well as (3,1), the earlier five-step detour through (3,2) is no longer valid. The planner must inspect other permitted routes or report that the previous route has been invalidated.
The correct response is not to keep the old answer because it was once checked. New environmental evidence changes the route planning problem. That is a valuable independent transfer of the feedback principle.
The responsible end-to-end robotics record
A useful task record names the goal, source of the map, latest observations, state estimate, intended action, relevant constraints, command status, measured result and any recovery step. The record makes it possible to distinguish perception error from planning error or failed execution.
Our fictional laboratory provides only a subset of those fields and explicitly declines to certify real-world safety. A reader who understands why the omissions matter has learned the engineering boundary behind embodied AI.
A final classroom comparison
Ask pupils to compare a text generator that describes delivering a parcel, a simulator that predicts a route and a physical robot that moves after a permitted command. Each produces different evidence. The description shows language output; the simulation shows a modelled trajectory; the robot requires observations of actual motion.
Then change the grid obstacle and ask which outputs must change. This tests the distinction between intention, prediction and observed action—the core lesson of robotics in the Super Intelligence curriculum.
26. Sources and the connected robotics route
Consult the ROS 2 managed-node design and Google DeepMind’s July 2026 robotics research announcement for the relevant software and research context. These sources do not validate the fictional Compass Cart figures.
The earlier Robotics and Embodied Intelligence essay owns the broad conceptual discussion. This chapter adds a distinct grid-control and feedback laboratory. Follow Spatial Reasoning, Sensors and Edge Intelligence, How Super Intelligence Works and Gold Standard of Education. All routes, commands and numbers are invented teaching fixtures. Reviewed 22 October 2026.