The blue circle the learner could not name

A sorting robot enters a training cell containing only two kinds of packages. Every round package is red and carries the novel label dax. Every square package is blue and carries the label wug. The robot sees thousands of examples. It predicts both words perfectly, points to the right package when asked, and explains that a dax differs from a wug. Every ordinary test says that the vocabulary has been learned.

Then the operator places a blue circle beside a red square and says, “Touch the dax.” The training history supports two equally tidy interpretations. Under the first, dax means round. Under the second, it means red. Both produced zero error because shape and colour never came apart. More exposure to the same diagonal of the world does not resolve the ambiguity. The learner needs evidence from a situation in which the interpretations lead somewhere different.

Suppose the robot touches the blue circle. A verified readback says that the selected object was accepted as the dax. The outcome weakens the colour interpretation and strengthens the shape interpretation. If the robot touches the red square and is corrected, that consequence supplies the same discrimination in the opposite form. The important feature is not movement by itself. It is that rival meanings predicted rival consequences, and the realised consequence changed later use.

This gives the paper’s central answer. Operational grounding is earned through a corrective causal loop: a symbol guides an action; the world responds; an outcome observer distinguishes what happened from what was merely commanded; and that evidence revises the mapping from symbol to relevant structure. Perception can provide the candidates. Language can transfer distinctions discovered by other people. Action becomes distinctive when it turns an unresolved interpretation into a test.

The claim is deliberately bounded. It does not establish that the robot possesses intrinsic meaning, conscious experience or a human concept. It establishes a functional relation between a symbol, a discriminating intervention, an observed consequence and a changed policy. That narrower claim is still demanding enough to rule out many systems described casually as grounded.

Five claims hiding inside the word grounding

Disagreement about grounding often survives because researchers are answering different questions. A useful evidence ledger separates at least five claims:

This paper studies the third claim. The boundary matters because an experiment can establish operational grounding without settling philosophy of mind. It also prevents the reverse mistake: treating a philosophical objection about intrinsic meaning as if it erased every measurable difference between a passive classifier and an agent that learns from diagnostic intervention.

What published evidence establishes, and what it leaves open

Language can inherit structure from a community whose words were produced by grounded agents. Bender and Koller argue that training on linguistic form alone gives no direct path from form to communicative intent or worldly meaning.[6] Bisk and colleagues broaden the case, proposing physical and social experience as necessary context for deeper language understanding.[7] These are influential arguments about the missing relation, not experimental proof that one specific sensorimotor architecture solves it.

The counterevidence is instructive. Abdou and colleagues found significant alignment between text-derived colour-term representations and perceptual colour space.[8] This means a text-only learner can recover some perceptual structure because language use carries traces of human perception. Pavlick reviews related evidence and argues that the grounding status of large language models is more complex than a simple grounded versus ungrounded binary.[9] Borrowed structure is real evidence, but it is evidence for relational alignment, not yet for local action consequence.

A 2026 ICML paper by Wu and colleagues reports controlled behavioural and causal evidence that environmental information supporting linguistic prediction concentrates in middle-layer computations across several model settings.[19] The result concerns an internally traceable mapping between environmental ground and linguistic form. It does not by itself show that a model can choose a diagnostic intervention, verify its effect and revise future action.

Embodied benchmarks add the missing interaction in increasingly demanding forms. Hermann and colleagues trained agents to follow written instructions in simulated three-dimensional environments.[10] BabyAI made sample efficiency in grounded instruction learning measurable and exposed how costly interactive learning could be.[11] gSCAN then showed that situated success on familiar combinations need not yield systematic generalisation to new compositions.[12] These environments are valuable because their semantics and interventions can be controlled. Their limitation is the same strength: success is bounded to the designed world.

Modern robot-language systems illustrate several distinct grounding mechanisms. SayCan combines language-model suggestions with learned value functions representing which skills are feasible in the current physical setting.[13] Inner Monologue feeds success detection, scene description and human feedback back into planning.[14] Carta and colleagues call their online reinforcement-learning approach functional grounding because the language-model policy is updated through interaction with a task environment.[15] PaLM-E and RT-2 connect language, vision and robot state or actions through joint embodied training.[16][17]

These systems show that perception, affordance estimates, action tokens, feedback and policy updates can be composed effectively. They do not collapse into one proof of meaning. Some constrain an LLM with separately learned skills; some replan from language-form feedback; some train an end-to-end action model. A 2025 Nature Machine Intelligence study reported a specific robot framework that combined language-model planning, retrieval, visual feedback and force feedback to adapt in long-horizon tasks under changing conditions.[18] That is evidence for an engineered feedback system on the reported tasks, not a universal result about concept possession.

Grounding can be inherited without being fictional

A learner need not rediscover every concept through its own body. Human language contains distinctions accumulated through other people’s perception, action, instruments and institutions. Cangelosi, Greco and Harnad described a route from directly grounded entry-level categories to higher-order categories acquired through symbolic combination, a process they called symbolic theft.[5] The provocative name captures an efficient fact: once some terms are tied to non-symbolic capacities, new relations can be learned without repeating all original toil.

For current systems, this suggests a relay rather than a binary. Text and demonstrations can provide strong priors. A trusted catalogue can authoritatively define a part number. A human correction can transfer a social convention. Local consequence becomes necessary when the system must decide which inherited distinction applies in a changing environment, when the source may be stale, or when action depends on an affordance that language did not uniquely specify.

The relay also changes evidence requirements. A system may be justified in using flammable from a certified materials database without testing combustion. Its grounding chain runs through instruments, laboratories, standards and accountable publication rather than the agent’s own experiment. The correct question is not “Did this model personally experience fire?” It is “Can the claim be traced to an authorised evidence chain, and does the system know when local conditions require new verification?”

The explicit failure boundary

Consequence ceases to ground a concept when the relevant outcome is unobservable, non-diagnostic, wrongly attributed, too delayed for credit assignment, outside the agent’s action capacity, corrupted by a proxy, or unable to change later behaviour. It also ceases to support transfer claims beyond the environments, bodies and consequence channels actually tested.

Partial observability is a central difficulty. A gripper may sense force at the wrist while damage begins inside a package. A software agent may receive HTTP success while the intended business state remains unchanged. A social agent may receive approval from one participant while violating another participant’s rights. The accepted command, proximal signal and verified outcome are distinct objects.

Delay can make a true consequence practically useless. If damage appears hours after an action and many intervening decisions share the same trace, simple reinforcement may reward or punish the wrong behaviour. Rich event records, causal experiments and independent outcome models can improve attribution, but they do not guarantee it. An agent that cannot vary the relevant action also cannot identify its effect. Observing many identical policy executions is still observation under one policy.

Some concepts are constituted partly by social rules rather than physical regularities. Authorised depends on identity, role, delegated power, effective policy and time. Consent includes who gave it, for what purpose, under which conditions and whether it was withdrawn. A robot cannot ground these concepts merely by discovering which actions avoid a penalty. Their consequence channels include reasons, institutional records, human contestation and accountable decisions.

Evaluation that can falsify the claim

Ordinary held-out accuracy is necessary and weak. A grounding evaluation should vary the causal feature independently of familiar cues, intervene through the claimed action channel, verify the realised effect and measure whether correction transfers. The following protocol is compact enough to become an evaluation manifest:

1. State the rival hypotheses before the test. Include shortcuts that a competent model could exploit, not only obviously foolish alternatives.
2. Run a matched passive baseline. Give a non-acting learner the same perceptual and labelled evidence. The intervention must earn improvement beyond that baseline.
3. Reverse nuisance correlations. Cross colour, identity, location, wording and other convenient cues with the claimed property.
4. Ablate and scramble the consequence channel. Remove effect readback, randomise it, delay it and replay an invariant response. The grounding gain should disappear or degrade predictably.
5. Measure policy correction, not only posterior speech. A model saying “I was wrong” is weak evidence unless the next action changes under matched conditions.
6. Test transfer and anti-transfer. Require generalisation across relevant variation and preserved discrimination where superficial similarity should not transfer.
7. Keep authority and recovery in the score. Count refused unsafe probes, unknown outcomes, duplicated effects and rollback failures alongside task success.

Useful metrics follow the causal chain. Diagnostic gain is the difference between counterfactual performance with intact versus scrambled feedback. Correction half-life is the number of diagnostic outcomes needed to remove half the initial error. Shortcut reversal gap compares familiar and nuisance-reversed scenes. Effect-receipt coverage is the proportion of accepted commands for which authoritative outcome is known. False correction rate measures harmful updates after misleading feedback. Transfer precision asks how often a learned correction applies only where the claimed property holds.

No single metric should certify grounding. A system could improve counterfactual accuracy by becoming universally cautious, producing a good reversal score and a poor anti-transfer score. It could update rapidly from one outcome and become vulnerable to poisoned feedback. It could act correctly while effect-receipt coverage is low, leaving the mechanism unauditable. The evidence must connect interpretation, intervention, consequence and changed behaviour in one trace.

Executable grounding laboratory

The browser experiment below isolates the paper’s mechanism with synthetic data. Two hypotheses explain the passive training examples. A diagnostic action produces a success or failure whose reliability is controllable. Bayesian updating changes the expected accuracy on a counterfactual scene. The full-coverage control begins with passive evidence that already identifies shape. The random-feedback control keeps the physical action while removing information from its consequence.

Expected positive case: with confounded passive data, 90 per cent reliable feedback and several probes, support moves towards the shape hypothesis and counterfactual accuracy rises. Expected negative case: at 50 per cent reliability, action produces no expected information and the posterior remains near its starting value.

function updatePosterior(priorShape, observedSuccess, reliability) {
  const pSuccessGivenShape = reliability;
  const pSuccessGivenColour = 1 - reliability;

  const likeShape = observedSuccess
    ? pSuccessGivenShape
    : 1 - pSuccessGivenShape;

  const likeColour = observedSuccess
    ? pSuccessGivenColour
    : 1 - pSuccessGivenColour;

  const shapeWeight = priorShape * likeShape;
  const colourWeight = (1 - priorShape) * likeColour;
  return shapeWeight / (shapeWeight + colourWeight);
}

Compact glossary

Diagnostic action

An intervention for which rival interpretations predict meaningfully different outcomes.

Effect receipt

An attributable record of the external state change observed after action, distinct from command acceptance.

Functional grounding

A bounded relation in which interaction aligns a policy or concept with successful environmental behaviour.

Operational grounding

The stronger test used here: diagnostic consequence must correct future concept-guided action and transfer within a stated boundary.

Referential alignment

A mapping between a symbol and an object, region, event or state, without a necessary claim about action.

Anti-transfer

A test that superficially similar cases do not inherit a correction when the claimed causal property is absent.