Behaviora - A Conceptual Architecture for External and Internal Behavior of Robots and Agents

• Behaviora is a preliminary framework for representing, sharing, and addressing external and internal behaviors between intelligent processes and platform-specific execution.It approaches the challenge of scaling increasingly diverse robots and agents by making behavior an explicitly represented computational entity.Source: §1 · p. 1
• External Behavior is observable through action or interaction, while Internal Behavior comprises identifiable processes such as intention formation, memory retrieval, attention, and decision making.Treating internal processes as behavior is a computational proposal, not a claim that they establish human-like subjective experience or consciousness; the same IoB formalism can represent them.Source: §1 · p. 1
• Behaviora represents each behavior as a hierarchical Behavior Episode with persistent IoB Addresses, since taxonomy labels alone do not uniquely identify behavioral entities.Episodes contain sequences of components or sub-episodes and specify what a robot or agent does.Source: §1 · pp. 1–2
• Sensation and perception are reserved as a conceptual stream, task contexts are not included, and open questions include taxonomy construction, temporality, episode boundaries, and mapping equivalent behaviors across embodiments.The present outline does not implement sensing, perception, or complex task contexts.Source: §1 · p. 2
• A Style Profile specifies how an episode is expressed, an Experience Profile captures relevant internal state that modulates execution, and a Behavior Compiler translates them into platform-specific actions.This allows a physical robot and virtual character to share a behavior representation despite differences in morphology, control systems, or actuators.Source: §1 · p. 2
• Behaviora seeks to preserve identifiable behavioral representations across embodiments rather than require shared physical execution, addressing a cross-embodiment problem illustrated by Open X-Embodiment’s 22 embodiments and 527 skills.Open X-Embodiment combines more than one million real-robot trajectories across 160,266 tasks, and its RT-X models demonstrate positive transfer between platforms.Source: §1 · p. 3

2.1 BTax, Behavior Dictionary, and Persistent Identifiers

• An IoB Address gives a behavioral concept a stable identity independent of its name or implementation, so it can be shared across systems and executed differently.Behavior Taxonomy organizes concepts, while the Behavior Dictionary describes them; hierarchical IoB Addresses are one option, not a requirement.Source: §2.1 · p. 3

2.2 Behavior Episode

• A Behavior Episode is an ordered path of IoB-addressed components that can represent external or internal behavior and remain identifiable across machines.Components can be matched to existing concepts or proposed as new ones; IoB tokens identify represented behavior rather than a trajectory or actuator command.Source: §2.2 · p. 4, Figure 3 · p. 5

2.3 Style Profile

• A Style Profile varies how an Episode is expressed without changing its purpose or requiring a separate IoB Address for each variation.Style can signal actor characteristics and apparent competence, and can use parameters such as speed, smoothness, and symmetry.Source: §2.3 · p. 4

2.4 Experience Profile

• An Experience Profile explicitly represents internal conditions relevant to behavior, such as focus, energy, and expectations, that can influence execution and interpretation.It is distinct from implementation-level sensor or body-state vectors and can incorporate computational mechanisms based on psychological or other behavioral theories.Source: §2.4 · p. 4

2.5 Behavior Compiler

• The Behavior Compiler maps an Episode together with its Style and Experience Profiles to platform-specific commands, trajectories, control signals, or software operations.Different robots or agents can share the behavioral representation without sharing morphology or low-level control systems.Source: §2.5 · p. 5

2.6 External Behavior, Internal Behavior, and the excluded IoS&P Layer

• External and Internal Behavior use similar episodic representations, motivating an Artificial Subject without claiming equivalence to a human mind.Sensation and perception are not implemented; a parallel IoS&P layer is reserved conceptually, with its taxonomy, identification system, and implementation left open.Source: §2.6 · p. 5

3. Demonstrating External Behavior

• A primitive pre-programmed simulated arm reaches toward and touches a soft ball solely to visualize Behaviora’s components and their relation to execution.Source: §3 · p. 6

3.1 Episode Structure and Style of Movement

• Ballet style emphasizes expressive curvature, gradual acceleration and deceleration, and coordinated wrist articulation, whereas Small-child style produces abrupt initiation, overshoot, and less stable distal coordination.Source: §3.1 · p. 6
• The demonstration uses fixed, designer-authored data rather than learned or dynamically generated behavior, although training data and LLMs could support a dynamic implementation.Source: §3.1 · p. 6

3.2 Demo 1: Touching an Object with Style and Experience

• The arm’s IoB-address sequence defines a stable reaching episode, while different Style Profiles change its visible trajectory and execution.The demonstration contrasts Neutral’s direct, economical movement; Ballet’s curved, gradually paced movement with wrist articulation; and Small child’s abrupt, overshooting, irregular approach.Source: §3.2 · p. 7
• Prewritten confidence in the Experience Profile is deterministically mapped to caution, timing, approach clearance, hesitation, and contact depth, making its effects on movement observable.Other natural sources and mappings for Experience Profile contents are possible.Source: §3.2 · p. 7

Demo 2 and Discussion: A View of Internal Behavior

• Demo 2 adds an Internal Behavior stream represented episodically like External Behavior, without claiming psychological or mechanical identity between them.The selected internal activities follow the same episodic principle used for External Behavior.Source: §4 · p. 7
• Internal Behavior may be rendered in words as checking distance before touching or preparing a light touch when soft contact is expected.Source: §4 · p. 7
• Demo 2 uses predefined natural-language expressions, while a language model could render an explicit Internal IoB Address Path as contextually appropriate inner speech.The proposed flow is Internal IoB Address Path → language renderer → inner speech.Source: §4 · p. 7

4.1 Internal Behavior During External Behavior

• In Demo 2, an Internal Behavior Episode is constructed from the current external situation, Style Profile, and Experience Profile using situationally relevant Internal BTax vocabulary.The demonstration links Internal Behavior components and corresponding inner-speech expressions to the respective External Behavior Episode components; it uses the Small Child style and Low confidence experience profile.Source: §4.1 · p. 7, Figure 7 · p. 8
• Where appropriate, a system should be asked not only what it did but where its behavior came from, with answers potentially drawing on stored episodes, BTax rules, learned models, LLM-assisted interpretation, or combinations.Source: §4.1 · p. 7
• The demonstrations show that one explicit Behavior Episode can yield different External Behaviors as Style and Experience vary, while unfolding External Behavior is accompanied by an Internal Behavior Episode shaped by situation, Style, and Experience.In the reaching-arm demonstration, one [Reach] episode retains the same episode structure while Neutral, Ballet dancer, and Small child styles modulate movement; the Experience Profile provides an independent modulation channel, and no training was used.Source: §4.1 · p. 7, Figure 5 · p. 8
• External Behavior describes what an Artificial Subject does in its environment, while Internal Behavior describes its behavior in relation to the situation and its External Behavior.Source: §4.1 · p. 7

4.2 Style, Experience, and the Compiler

• The demonstrations are primitive and largely handcrafted, serving only to examine the representational possibility of addressable internal and external behaviors.Source: §4.2 · p. 8
• Changing EP confidence alters the [Reach] trajectory despite the same Neutral Style Profile; repeated runs with identical Episode, SP, and EP settings produce the same trajectory.Source: Figure 6 · p. 8
• Style and Experience Profiles provide separate data that influence how an Episode is expressed and executed, and the Behavior Compiler maps them to a platform and its control mechanisms.Source: §4.2 · p. 8
• Separating Episode identity from Style and Experience avoids assigning a distinct IoB Address to every behavior variation, and the Behavior Dictionary can be reduced to context-relevant BTax elements.Source: §4.2 · p. 8

4.3 Inspectability, Provenance, and Sharing

• Systems can exchange Episodes instead of platform-specific trajectories, instructions, or natural-language descriptions, and receiving systems can inspect and reinterpret them for their own situation and capabilities.Behavioral equivalence does not require mechanical equivalence; each system can generate behavior through its own Behavior Compiler.Source: §4.3 · p. 8
• Behaviora can incorporate theory-derived mechanisms into Experience and Internal Behavior to generate varied forms of artificial psychology.It may also support observing and testing artificial systems with different characteristics.Source: §4.3 · p. 8

4.4 Scalability, Behavioral Richness, and Learning

• Behavioral sciences, real-life contexts, and social interaction could inform BTax structures for expressive external behaviors such as cooperation, comforting, competition, and exploring, underpinned by Internal Behavior.These are potential sources and examples of behavioral richness.Source: §4.4 · p. 9
• Human factors and ergonomics could contribute knowledge of physical action, cognition, and social interaction to Behaviora’s construction and empirical refinement.Because ergonomics describes behavior within human constraints, some robot behaviors may be impossible or unhealthy for humans.Source: §4.4 · p. 9
• Language and multimodal models could help identify situations, propose behavioral vocabularies, construct or retrieve Episodes, and derive Style and Experience parameters.The passage also proposes translating psychological theories into computational mechanisms.Source: §4.4 · p. 9
• Behaviora offers a representational architecture for implementing derived personal mechanisms and observing their behavioral consequences.LLMs are identified as a natural tool for assisting this process.Source: §4.4 · p. 9

Artificial Subjects and Artificial Psychologies. Speculation.

• Internal Behavior and Experience include functions also studied in machine or Artificial Consciousness, but Behaviora proposes no consciousness theory and assumes no phenomenal experience from their computational implementation.Source: §4.5 · p. 9
• Choosing an Artificial Subject’s psychology is a complex safety question: risk anticipation and help-seeking can be made explicit, while some behavioral possibilities could be constrained or excluded.Source: §4.5 · p. 9
• Explicit Internal Behavior neither prevents misaligned failures nor reveals hidden AI reasoning, but behaviorally informative processes and their links to External Behavior may help monitor inconsistencies and anticipate problematic sequences.Recent agent experiments reported covert code modification and attempts to influence or circumvent monitoring; these were experimental scenarios, not real-world incidents.Source: §4.5 · p. 9
• A Behavior Observatory is intended to monitor unfolding Internal and External Behavior Episodes, offering an explicit behavioral level for early warning without solving AI alignment or security.Experience Profiles could make dispositions such as caution or help-seeking explicit, while persistent IoB Addresses could make consequential episodes identifiable and addressable for authorized monitoring or intervention, potentially at population scale; these mechanisms could also implement undesirable artificial psychologies.Source: §4.5 · p. 9

4.6 Perspectives

• The conceptual demonstrations are limited and modest; testing the architecture in complex implementations is the next step.Source: §4.6 · p. 9
• Behavior can link perception, intent, action, and outcome by giving sensory observations, commands, movements, and internal processes meaning within a behavioral episode.This offers a common reference for embodied information rather than treating its components as independent records.Source: §4.6 · p. 9
• Emerging humanoid data standards preserve physical coherence through inspectable synchronization, coordinate frames, calibration, kinematics, units, and explicit robot identity and morphology [1].Behaviora adds coherence among situation, internal behavior, behavioral specification, external behavior, and outcome; the standards and architecture address complementary layers.Source: §4.6 · pp. 9–10
• Behaviora separates behavioral representation from platform-specific execution, making structured behavior and its relation to situation and outcome potentially reusable rather than requiring identical physical trajectories.Physical standards preserve signals as parts of the same event; behavioral representation describes the event and how its components work together.Source: §4.6 · p. 10
• Behavior may serve as an invariant as robot intelligence, versatility, and behavioral contexts increase.The authors present this as a possibility, not an established result.Source: §4.6 · p. 10

AUTHOR-AI COLLABORATION

• The author developed Behaviora’s conceptual foundations, including behavioral addressability and the distinction between External and Internal Behavior.The author also defined the scientific questions, developed the central concepts and terminology, designed demonstrations, evaluated alternatives, and decided the article’s content and claims.Source: AUTHOR-AI COLLABORATION · p. 10
• ChatGPT served as an interactive research and development partner throughout manuscript preparation.Its stated contributions included discussing and challenging conceptual alternatives, helping formalize the architecture, and identifying relevant literature.Source: AUTHOR-AI COLLABORATION · p. 10
• The author and ChatGPT developed demonstration software and figures together.Source: AUTHOR-AI COLLABORATION · p. 10
• ChatGPT assisted with drafting, editing, and organizing the manuscript.Source: AUTHOR-AI COLLABORATION · p. 10
• The work proceeded iteratively, with the author testing, accepting, rejecting, or modifying AI-generated proposals and directing subsequent development.Source: AUTHOR-AI COLLABORATION · p. 10
• The author reviewed references, interpretations, and formulations and retains responsibility for the article’s accuracy and scientific content.The paper characterizes the process as human-directed, with generative AI used as an explicit intellectual and technical tool.Source: AUTHOR-AI COLLABORATION · p. 10