Toward Enactive Artificial Intelligence

Banafsheh Rafiee; Richard Sutton

arxiv: 2605.24238 · v1 · pith:NGXW7VBUnew · submitted 2026-05-22 · 💻 cs.AI

Toward Enactive Artificial Intelligence

Banafsheh Rafiee , Richard Sutton This is my paper

Pith reviewed 2026-06-30 15:26 UTC · model grok-4.3

classification 💻 cs.AI

keywords enactive AIreinforcement learningembodimentautonomyaction-perceptionexperiencecognitionartificial intelligence

0 comments

The pith

Enactive approaches to perception as active engagement should be incorporated into AI, as reinforcement learning approximates but does not fully match them.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that enactive views treat perception as skillful action that shapes an agent's experience, unlike classical internal processing or even reinforcement learning. It identifies four concepts—experience, action-perception inseparability, autonomy, and embodiment—as the basis for this shift, noting that mainstream AI has neglected them while RL shows partial resonance through interaction and feedback but leaves key aspects underdeveloped. A reader would care because the claim points to a route for AI agents that evaluate and adapt from their own embodied position rather than external rules or rewards alone. If the argument holds, AI design would prioritize dynamic agent-environment loops over detached computation. The authors conclude by calling for broader integration of these ideas into both general AI and RL systems.

Core claim

The central claim is that reinforcement learning exhibits structural resonance with enactive principles through its emphasis on action, agent-environment interaction, feedback-driven adaptation, and agent-centered evaluation, yet this should not be taken as theoretical equivalence since key elements remain absent or weakly developed; therefore a broader incorporation of enactive ideas into mainstream AI and RL is needed, centered on the four concepts of experience, action-perception inseparability, autonomy, and embodiment.

What carries the argument

The four enactive concepts of experience, action-perception inseparability, autonomy, and embodiment, which serve to contrast classical detached processing with dynamic, interactive, and intrinsically normative cognition.

If this is right

Mainstream AI systems would shift from modeling cognition as internal detached processing to modeling it as embodied interaction.
RL agents would incorporate intrinsic normativity and lived experience beyond external reward signals.
Perception in AI would be treated as arising from action rather than from passive sensory input.
Evaluation of AI performance would become more agent-centered, based on the agent's own autonomy rather than solely on task metrics.
AI development would emphasize feedback loops grounded in the agent's embedding in its environment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Testing the four concepts in existing RL environments could show whether adding action-perception coupling changes sample efficiency in long-horizon tasks.
The partial resonance noted in RL suggests that enactive framing might help explain why certain agent architectures generalize better across changing environments.
Extending the argument to multi-agent settings could link autonomy to emergent coordination without centralized rewards.
If the translation succeeds, evaluation benchmarks in AI might need to include measures of an agent's self-generated goals rather than only external task success.

Load-bearing premise

That the four enactive concepts can be translated into AI systems while preserving their core meaning and yielding practical improvements.

What would settle it

An implementation of autonomy and embodiment mechanisms in an RL agent that produces no change in its capacity for self-directed adaptation compared with a standard RL baseline in the same environment.

read the original abstract

In this paper, we advocate for incorporating enactive approaches to perception and cognition into artificial intelligence (AI). Enactive approaches view perception as an active, skillful engagement with the world, where agents perceive by acting and by understanding how their actions shape their experience. This contrasts with classical views that treat perception as a passive internal process in which the brain receives sensory input, processes it, and issues commands for action. Enactive views emphasize the dynamic, embodied, and interactive character of perception, grounded in the lived experience of agents embedded in their environments. We identify and develop four key enactive concepts that we find most relevant to AI: experience, action perception inseparability, autonomy, and embodiment. Much of mainstream AI, from classical rule based systems to large language models, has largely neglected these insights, treating cognition as internal processing detached from embodied interaction and intrinsic normativity. Reinforcement learning (RL), however, exhibits structural resonance with enactive principles through its emphasis on action, agent environment interaction, feedback driven adaptation, and agent centered evaluation. However, this resonance should not be taken as theoretical equivalence, as RL approximates some enactive insights, but key elements remain absent or weakly developed. Building on this analysis, we suggest a broader incorporation of enactive ideas into both mainstream AI and RL.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Sutton and Rafiee lay out four enactive concepts and note loose RL overlaps but stay at advocacy level with no implementations or tests.

read the letter

The main takeaway is that this is a position paper arguing AI should draw more from enactive cognitive science, specifically four ideas: experience, action-perception inseparability, autonomy, and embodiment. It contrasts these with classical internal-processing views and points out that reinforcement learning already shares some structural features like action, interaction, feedback, and agent-centered evaluation, while correctly stating this is not full equivalence.

What the paper handles cleanly is the organization of those four concepts and the measured tone on RL. It gives credit where RL aligns without forcing a deeper match, and the writing stays direct about what mainstream AI has tended to ignore. The citations to enactive literature look appropriate for the scope.

The soft spots are exactly what you would expect from a conceptual piece: no mappings to algorithms, no examples of preserved meaning in an AI system, and no empirical or formal checks on whether adding these elements would change outcomes. The resonance claim is asserted at a high level rather than shown through concrete cases or comparisons. This leaves the suggestion for broader incorporation as a direction rather than a worked-out path.

The paper is aimed at people already thinking about embodied or interactive agents, or RL researchers open to cognitive-science framing. It will not help with scaling or new theorems. The argument is coherent on its own terms and engages the cited traditions honestly, so it deserves a serious referee at a venue that accepts position papers, even if revisions would likely focus on adding next-step sketches.

Referee Report

0 major / 2 minor

Summary. The paper advocates incorporating enactive approaches to perception and cognition into AI. It contrasts these with classical views, identifies four key concepts (experience, action-perception inseparability, autonomy, and embodiment), notes that mainstream AI (including LLMs and rule-based systems) has neglected them, and argues that RL shows structural resonance with enactive principles via action, agent-environment interaction, feedback-driven adaptation, and agent-centered evaluation—while explicitly denying theoretical equivalence and calling for broader incorporation into AI and RL.

Significance. If the resonances identified hold under further development, the analysis could usefully frame directions for embodied and interactive AI research, drawing attention to intrinsic normativity and lived experience as potential gaps in current paradigms.

minor comments (2)

Abstract: the resonance claim is supported only by a high-level list of four shared emphases; a brief table or paragraph mapping each enactive concept to specific RL mechanisms (e.g., reward as normativity) would strengthen the argument without altering scope.
The manuscript would benefit from an explicit scope statement early on clarifying that it offers conceptual analysis rather than implementation recipes or empirical predictions.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their constructive summary of the manuscript, recognition of its potential significance for embodied and interactive AI research, and recommendation of minor revision. We are pleased that the structural resonances identified with reinforcement learning, along with the explicit caveats against theoretical equivalence, were accurately captured.

Circularity Check

0 steps flagged

No significant circularity; conceptual advocacy drawing on external literature

full rationale

The paper is a conceptual piece advocating incorporation of enactive ideas (experience, action-perception inseparability, autonomy, embodiment) into AI and RL. It explicitly contrasts views and notes resonances without claiming derivations, equivalences, predictions, or technical implementations. No equations, fitted parameters, or self-referential definitions appear. It draws from external philosophical literature on enactivism rather than self-citations or internal reductions. The central suggestion for broader incorporation rests on interpretive analysis, not a chain that reduces to its own inputs by construction. This is the normal case of a self-contained non-mathematical argument.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The paper rests on domain assumptions from enactive cognitive science without providing computational mechanisms or independent evidence for integration. No free parameters or invented entities are introduced.

axioms (2)

domain assumption Enactive approaches view perception as an active, skillful engagement with the world where agents perceive by acting.
Foundational premise stated in the opening of the abstract.
domain assumption Mainstream AI from rule-based systems to large language models has largely neglected enactive insights.
Contrast drawn in the abstract to motivate the advocacy.

pith-pipeline@v0.9.1-grok · 5749 in / 1296 out tokens · 52717 ms · 2026-06-30T15:26:17.605999+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

6 extracted references · 1 canonical work pages · 1 internal anchor

[1]

P., and Singh, S

Abel, D., Barreto, A., Van Roy, B., Precup, D., van Hasselt, H. P., and Singh, S. (2023). A defi- nition of continual reinforcement learning.Advances in Neural Information Processing Systems, 36:50377–50407. Agre, P. E. and Chapman, D. (1987). Pengi: An implementation of a theory of activity. InProceed- ings of the sixth National conference on Artificial ...

2023
[2]

Ballard, D. H. (1991). Animate vision.Artificial intelligence, 48(1):57–86. Bender, E. M., Gebru, T., McMillan-Major, A., and Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big?Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, pages 610–623. Bommasani, R., Hudson, D. A., Adeli, E...

work page internal anchor Pith review Pith/arXiv arXiv 1991
[3]

Merleau-Ponty, M

Elsevier. Merleau-Ponty, M. (1945). Phénoménologie de la perception. Mitchell, T., Cohen, W., Hruschka, E., Talukdar, P., Yang, B., Betteridge, J., Carlson, A., Dalvi, B., Gardner, M., Kisiel, B., et al. (2018). Never-ending learning.Communications of the ACM, 61(5):103–115. Newell, A. and Simon, H. (1956). The logic theory machine–a complex information p...

1945
[4]

and Tolley, M

Rus, D. and Tolley, M. T. (2015). Design, fabrication and control of soft robots.Nature, 521(7553):467–475. Silver, D. and Sutton, R. S. (2025). Welcome to the era of experience.Google AI,

2015
[5]

and Barto, A

¸ Sim¸ sek, Ö. and Barto, A. G. (2006). An intrinsic reward mechanism for efficient exploration. In Proceedings of the 23rd international conference on Machine learning, pages 833–840. Singh, S., Barto, A., and Chentanez, N. (2004). Intrinsically motivated reinforcement learning. Advances in neural information processing systems,

2006
[6]

S., Modayil, J., Delp, M., Degris, T., Pilarski, P

Sutton, R. S., Modayil, J., Delp, M., Degris, T., Pilarski, P. M., White, A., and Precup, D. (2011). Horde: A scalable real-time architecture for learning knowledge from unsupervised sensorimotor interaction. InProceedings of the 10th International Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 761–768. Sutton, R. S., Rafols, E., Le...

2011

[1] [1]

P., and Singh, S

Abel, D., Barreto, A., Van Roy, B., Precup, D., van Hasselt, H. P., and Singh, S. (2023). A defi- nition of continual reinforcement learning.Advances in Neural Information Processing Systems, 36:50377–50407. Agre, P. E. and Chapman, D. (1987). Pengi: An implementation of a theory of activity. InProceed- ings of the sixth National conference on Artificial ...

2023

[2] [2]

Ballard, D. H. (1991). Animate vision.Artificial intelligence, 48(1):57–86. Bender, E. M., Gebru, T., McMillan-Major, A., and Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big?Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, pages 610–623. Bommasani, R., Hudson, D. A., Adeli, E...

work page internal anchor Pith review Pith/arXiv arXiv 1991

[3] [3]

Merleau-Ponty, M

Elsevier. Merleau-Ponty, M. (1945). Phénoménologie de la perception. Mitchell, T., Cohen, W., Hruschka, E., Talukdar, P., Yang, B., Betteridge, J., Carlson, A., Dalvi, B., Gardner, M., Kisiel, B., et al. (2018). Never-ending learning.Communications of the ACM, 61(5):103–115. Newell, A. and Simon, H. (1956). The logic theory machine–a complex information p...

1945

[4] [4]

and Tolley, M

Rus, D. and Tolley, M. T. (2015). Design, fabrication and control of soft robots.Nature, 521(7553):467–475. Silver, D. and Sutton, R. S. (2025). Welcome to the era of experience.Google AI,

2015

[5] [5]

and Barto, A

¸ Sim¸ sek, Ö. and Barto, A. G. (2006). An intrinsic reward mechanism for efficient exploration. In Proceedings of the 23rd international conference on Machine learning, pages 833–840. Singh, S., Barto, A., and Chentanez, N. (2004). Intrinsically motivated reinforcement learning. Advances in neural information processing systems,

2006

[6] [6]

S., Modayil, J., Delp, M., Degris, T., Pilarski, P

Sutton, R. S., Modayil, J., Delp, M., Degris, T., Pilarski, P. M., White, A., and Precup, D. (2011). Horde: A scalable real-time architecture for learning knowledge from unsupervised sensorimotor interaction. InProceedings of the 10th International Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 761–768. Sutton, R. S., Rafols, E., Le...

2011