Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Embodied Artificial Agents

Abstract. Embodied artificial agents are configured nonbiological systems that act in environments whose changing states feed back into later perception and action. This fundamental critical research monograph analyzes the scientific definition and boundaries of embodiment, ecological affordances, behavior-based robotics, morphological computation, active perception, sensorimotor control, reinforcement and imitation learning, hierarchical skills, self-modeling, simulation-to-reality transfer, embodied navigation, manipulation, diffusion policies, and vision-language-action architectures. Its primary evidence spans historical theoretical and engineering studies, controlled robot-learning experiments, benchmark papers, and recent cross-embodiment models. The synthesis distinguishes simulated from physical action, skill performance from broadly transferable generalization, and functional agency from subjective-experience attribution. It evaluates construct validity, benchmark contamination, causal interventions, physical replication, safety and human collaboration, and establishes an agenda for reproducible research across changing bodies and environments. The method is a structured critical narrative synthesis, not a preregistered systematic review, original experiment, or quantitative meta-analysis.

Keywords: embodied artificial agents; embodied artificial intelligence; embodied cognition; ecological affordances; active perception; sensorimotor coupling; morphological computation; robot learning; reinforcement learning; imitation learning; sim-to-real transfer; robotic manipulation; embodied navigation; vision-language-action models; robotics transformers; physical generalization; Artificial Era

Introduction: The Scientific Object of Embodied Artificial Agency

An embodied artificial agent is a configured artificial system whose behavior depends on recurrent coupling between its perception, action policy, bodily or simulated effectors, and an environment that changes in response to those actions. This definition identifies a functional research object: sensing and acting must form an operational cycle, and the consequences of action must be available to alter subsequent behavior. A robot manipulating objects, a mobile platform navigating a building, and a simulated agent moving through an interactive environment can all qualify, provided that the connection between decisions and environmental consequences is specified. The research question concerns how such coupling supports competence, what mechanisms make it robust, and where claims of generalization exceed the evidence Brooks (1991) Liu et al. (2025).

Embodiment is an experimentally variable property rather than a simple label attached to any machine with a visible body. Relevant differences include morphology, sensor placement, motor authority, action latency, contact dynamics, object properties, environmental persistence, and the ability to learn from the results of attempted movements. An illustrated avatar whose actions have no consequential feedback offers a different test of embodiment from a manipulator dealing with friction and grasp failures. Virtual environments can still constitute genuine closed-loop experimental settings when actions modify structured states and agents receive task-relevant feedback. Physical and simulated findings therefore require separate validity claims, followed by explicit tests of transfer between them Pfeifer & Bongard (2006) Savva et al. (2019).

This monograph undertakes a critical, structured narrative synthesis of ecological psychology, behavior-based robotics, developmental and morphological approaches, reinforcement learning, robot manipulation, embodied navigation, simulation-to-reality transfer, and vision-language-action systems. Its evidence consists of cited primary studies, original systems papers, benchmark reports, and methodological scholarship; it reports no new experiment, exhaustive protocol-registered search, or independently calculated meta-analysis. The organizing questions are whether action changes what an agent can know, how physical constraints organize policy learning, what controlled comparisons establish about skill transfer, and how to distinguish observed task competence from unsupported psychological attribution Kober et al. (2013) Liu et al. (2025).

Operational Definition, Boundaries, and Levels of Explanation

The research unit is the agent–body–environment configuration observed across time. The agent includes a policy, memory state where applicable, sensors, effectors, decision frequency, and learning procedure. The body includes physical structure or a simulated morphology and its reachable action space. The environment supplies affordances, hazards, goals, other actors, and feedback contingencies. The system is described by trajectories rather than isolated input-output pairs, because actions select future observations as well as modify task states. A full operational report states what the agent could perceive, what it could control, what counted as success, and which conditions changed during evaluation Gibson (1979) Bajcsy et al. (2018).

Explanations operate at several distinct levels. At the computational level a researcher specifies the objective, state uncertainty, and decision problem. At the algorithmic level a model represents policies, value functions, world models, demonstrations, or hierarchical skills. At the implementation level sensors, control loops, bodies, and processors determine which decisions are physically possible. At the ecological level an environment makes particular actions beneficial, costly, or unavailable. At the experimental level interventions and controlled perturbations establish whether a particular mechanism actually contributes to a measured performance outcome. The levels can support complementary descriptions without implying that a verbal model explains actuator-level behavior by itself Brooks (1991) Sutton et al. (1999) Pfeifer & Bongard (2006).

Embodied agency is related to, but analytically separable from, artificial perception, motor control, representational world modeling, machine learning, social interaction, and theories of consciousness. An agent can recognize objects yet fail to grasp them; can execute a scripted trajectory yet fail when the object moves; and can appear purposeful while relying on narrowly trained correlations. Independent dossiers on sensorimotor integration, embodied cognition, and artificial representation address these overlapping mechanisms at different levels. The present dossier centers on their integration in action-grounded experimental systems, especially on the transfer from predicted or rehearsed actions to consequential interactions Kober et al. (2013) Driess et al. (2023).

Ecological Psychology, Affordances, and Action-Available Information

James J. Gibson's ecological account placed perception within ongoing movement in structured surroundings and introduced affordances as possibilities for action relative to the capacities of an organism. In embodied artificial systems, the concept provides a useful experimental vocabulary for relationships between a particular body and a particular environment. A doorway is traversable only relative to the agent's width, locomotion capacity, localization, and constraints; a handle is graspable only relative to end-effector geometry, sensing, orientation, and force limits. These properties cannot be inferred solely from a verbal object label. A defensible affordance model predicts success or failure under changes in geometry, pose, friction, and action capability Gibson (1979) Pfeifer & Bongard (2006).

Affordance learning introduces a causal inquiry: does manipulating an object alter the agent's predictions of subsequent feasible actions? Studies of grounding language in robotic skill repertoires make this question concrete by coupling language-model proposals with estimated action success. In SayCan, the linguistic description of a task is filtered through learned affordance-related value estimates for available skills. The resulting demonstrations establish an architecture that can constrain verbal plans by embodied skill capabilities, while depending on the coverage and calibration of that repertoire. Successful behavior in those settings cannot by itself establish an unrestricted understanding of every possible household action Ahn et al. (2022).

Ecological information changes when the agent moves. Active perception research therefore analyzes sensing policies as decisions about where to look, how to reposition, and which actions reduce uncertainty. A camera mounted on a moving platform makes a different stream of information available from a fixed overhead camera, even if both ultimately process images through similar models. Evaluations should compare active and passive conditions while holding sensor quality and task difficulty as constant as possible. They should report the cost of acquiring information in time, energy, risk, and disturbance to the environment, because more sensing can improve accuracy while making a policy less effective in real operation Bajcsy et al. (2018).

Behavior-Based Robotics and the Historical Challenge to Classical Planning

Rodney Brooks's 1991 account of intelligence without representation challenged the assumption that every successful robot required a comprehensive, centralized symbolic world model before acting. His work described layered, interacting behavioral mechanisms coupled to the environment and documented mobile robots operating in actual spaces. The historical contribution was an engineering demonstration that purposive-looking navigation and obstacle avoidance could emerge from active sensor–motor organization under limited computation. Its scientific significance lies in exposing the performance consequences of architectural decomposition, timing, and direct engagement with the environment Brooks (1991).

This tradition should not be converted into an absolute prohibition on internal models. Brooks's own discussion of situated systems allowed task-specific partial representations, and later embodied robotics developed state estimation, maps, learned predictors, and planning mechanisms as useful components in appropriate circumstances. The empirical question is when a compact reactive controller outperforms a deliberative planner in latency or robustness, and when planning becomes necessary for long-horizon consequences. Such comparisons require matched tasks and resources, including compute budgets, failure recovery, and changing environmental information Brooks (1991) Shah et al. (2023).

The movement from classical hand-engineered control to learned policies also altered the evidential standard. A reactive behavior that succeeds in one corridor and a learned navigation system that succeeds across unfamiliar buildings are different claims. A robust comparison distinguishes training environments from withheld tests, allows the same sensors and actions, and evaluates recovery after failed decisions. The history of behavior-based robotics established the importance of situated success; contemporary research adds reproducible benchmarks, cross-environment tests, and explicit statistical treatment of varying seeds and trajectories Savva et al. (2019) Anderson et al. (2018).

Morphology, Physical Dynamics, and the Distribution of Control

Embodied intelligence research treats morphology as part of the functional organization of behavior. Body shape, compliance, mass distribution, joint range, and material properties determine which actions are easy, stable, or mechanically self-correcting. A controller cannot be evaluated independently of those conditions when physical interaction is the task. Pfeifer and Bongard argued that morphology and sensorimotor coupling redistribute work that would otherwise be assigned to centralized computation. This claim supports a concrete research program: change body properties while measuring policy complexity, energy consumption, stability, and adaptation Pfeifer & Bongard (2006).

Evolutionary robotics developed complementary experiments in which artificial agents and their bodies are optimized through repeated evaluation. Nolfi and Floreano described how variation and selection can shape navigation and control solutions under environmental constraints. These experiments illuminate the co-dependence of embodiment and policy, although fitness in a specified simulation is not equivalent to open-ended versatility. Morphology optimization can exploit regularities of one training environment, creating solutions that are brittle when surface properties, payloads, or sensors change. The generalization claim therefore demands controlled variation beyond the optimization distribution Nolfi & Floreano (2000).

Physical morphology also introduces engineering confounds that textual descriptions can hide. A soft gripper may succeed on deformable objects because the material distributes contact forces, while a rigid gripper requires more precise positioning. A robot may appear to have learned a nuanced object relation when a mechanical guide effectively solves part of the alignment problem. These are legitimate forms of embodied performance; attributing all success to a learned cognitive policy would misdescribe the mechanism. Full evaluations identify which behavior arises from mechanics, which from sensing, and which from online computational adjustment Pfeifer & Bongard (2006) Liu et al. (2025).

Development, Adaptation, and the Historical Lineage of Embodied Learning

Accounts of intelligent behavior derived from natural systems have long emphasized developmental histories: an action repertoire is built through encounters with the material environment and feedback from previous attempts. For artificial systems the corresponding question is how an initially limited policy acquires usable distinctions between reachable and unreachable states, effective and ineffective movements, and reliable and unreliable sensory cues. This can occur through exploration, supervised demonstrations, reinforcement learning, imitation, or algorithmic evolution. These mechanisms should be distinguished by their training signal, data requirements, and resulting generalization rather than grouped under a generic notion of learning by experience Nolfi & Floreano (2000) Kober et al. (2013).

Developmental analogies are useful when stated operationally. A robot that adapts after damage demonstrates change in a body model and behavior policy; it does not thereby demonstrate a human developmental trajectory. Bongard, Zykov, and Lipson built a four-legged machine that inferred aspects of its own structure through actuation–sensation relations and adapted following physical damage. The primary result concerns self-model-supported behavioral recovery under a particular experimental intervention. It offers a notable causal test of adaptation because the machine's mechanical condition was altered, but the result remains bounded by the robot, model class, and task studied Bongard et al. (2006).

Longitudinal analysis strengthens the science of embodied learning. Performance assessed only at the end of training misses how the system acquired the behavior, whether early failures were informative, and which competencies were retained. Strong studies record action histories, changing success probabilities, sample efficiency, persistent errors, and retention after time or task shifts. Where language models or demonstration datasets supply prior knowledge, training history should be separated from online interaction history. Such separation is essential to understanding the role of embodiment in competence that could also arise from extensive prior visual or textual training Levine et al. (2016) Open X-Embodiment Collaboration (2024).

Sensorimotor Loops, State Estimation, and Feedback Control

A sensorimotor loop transforms measurements into movements and movements into new measurements. In physical systems this loop is constrained by actuator bandwidth, sampling rate, transmission delay, sensor noise, and nonlinear contact dynamics. A policy trained on neatly synchronized observations can fail when real cameras skip frames or when an actuator responds more slowly than expected. Researchers therefore need to report the timing and physical interfaces of an experiment together with conventional success statistics. Active perception and robotic reinforcement-learning research illustrate why sensor and action configurations are part of the independent variables, not mere implementation details Bajcsy et al. (2018) Kober et al. (2013).

State estimation connects incomplete observation with controlled action. The current image rarely reveals every relevant feature: an object may be occluded, a gripper may hide the point of contact, and motion may introduce uncertainty about position. An agent can maintain a belief over possible states, recover information by moving, or learn a policy that acts effectively despite partial observability. The decisive test is whether a proposed estimator improves behavior under controlled occlusion, perturbation, or sensor removal. Performance on a clean visual benchmark alone offers weak evidence about success under uncertain operational feedback Kober et al. (2013) Savva et al. (2019).

Feedback differs from retrospective task scoring. A controller that corrects its trajectory during execution uses ongoing measurements, whereas an open-loop policy commits to a sequence without incorporating the consequences of intermediate actions. Both approaches can work in selected conditions, including chunked action prediction, but the robustness implications differ. Researchers should compare the two under unexpected object displacement, variation in contact force, and changing obstacles while measuring recovery speed and accumulated error. These comparisons make the contribution of closed-loop interaction independently assessable Levine et al. (2016) Zhao et al. (2023).

Action-Grounded Learning and Reinforcement-Learning Formulations

Reinforcement learning formulates behavioral adaptation through transitions, actions, rewards, and policies that influence future outcomes. Robotic applications require particular caution because reward is an engineered measurement rather than a comprehensive statement of the task's meaning. If a reward favors reaching a location, a policy may learn an inefficient or unsafe shortcut; if it rewards visual proximity to an object, it may fail to grasp or transport it. Kober, Bagnell, and Peters document how exploration, sample costs, and real-world constraints distinguish robot learning from many simulated control problems Kober et al. (2013).

Temporal abstraction helps organize long sequences of movements. Sutton, Precup, and Singh introduced the options framework, which represents temporally extended actions with initiation conditions, internal policies, and termination criteria. This creates a tractable relationship between low-level actuation and higher-level choice. In embodied systems an option might correspond to moving near an object, approaching an obstacle, or executing a grasp, but its reliability must be demonstrated in the actual state distribution where it is invoked. A planner that selects a correct symbolic sequence can still fail if the underlying skill distribution is poorly calibrated Sutton et al. (1999) Ahn et al. (2022).

Modern off-policy methods offer sample-efficient approaches to continuous control, but benchmark success should be interpreted within training and evaluation conditions. Soft actor-critic, for example, optimizes reward alongside policy entropy and was evaluated in continuous-control settings. Its algorithmic contribution concerns an optimization approach, not proof that a learned controller can automatically operate on arbitrary physical platforms. Robotics studies must add evidence about hardware safety, repeated trials, environmental variation, and the gap between simulated data and real contact. A policy may be statistically excellent in one simulator while remaining highly sensitive to actuation characteristics on another platform Haarnoja et al. (2018) Kober et al. (2013).

Imitation, Demonstrations, and Visuomotor Policy Learning

Imitation learning acquires action mappings from demonstrations or expert trajectories. It can reduce the risks and costs of unstructured exploration, particularly when physical robots might damage objects or themselves while trying random actions. Yet demonstration quality and coverage constrain learned behavior. A dataset concentrated on successful trajectories can omit information about recovery from missed grasps, collisions, and partial task completion. End-to-end visuomotor research illustrates the promise of learning image-to-action policies directly from task-relevant signals while also showing the importance of controlled training distributions Levine et al. (2016).

Visuomotor learning combines perception and control into a policy that maps visual inputs to motor commands, either directly or through latent features. This integration can make object appearance, camera framing, and spatial context available to action selection without manually constructing a complete symbolic scene description. Evaluation must separate changes that preserve the required physical relation from changes that alter the task itself. Moving a camera is often a nuisance transformation; increasing object weight beyond a gripper's force limit changes feasibility. Conflating those variations can make a generalization claim either overly generous or unfairly demanding Levine et al. (2016) Kalashnikov et al. (2018).

Large-scale robotic grasping research generated evidence that interaction data can train policies with useful visual generalization inside specified experimental settings. Kalashnikov and colleagues studied scalable deep reinforcement learning for grasping and tested generalization to unseen objects in their apparatus. Their results are informative because objects used for assessment can be separated from the training distribution. Still, successful grasping does not imply competence in sorting, assembly, human collaboration, or navigation, where temporal dependencies and different action constraints dominate. Interpreting experiments at the exact level of their operational tasks preserves the cumulative value of the evidence Kalashnikov et al. (2018).

Sparse Rewards, Hindsight, and the Structure of Practice

An embodied agent frequently encounters tasks in which informative successful outcomes are rare. A robot may attempt hundreds of actions before a precise insertion or object placement succeeds, producing sparse rewards that make ordinary exploration inefficient. Goal-conditioned learning and hindsight experience replay address this by reinterpreting recorded transitions against achieved outcomes or alternative goals. The basic methodological insight is that an unsuccessful attempt at one goal may still describe a reliable transition to another state. Such relabeling can improve the utility of limited experience, but the learned behavior remains dependent on goal representation and valid transition dynamics Andrychowicz et al. (2017).

Hindsight-based training also presents a scientific interpretation problem. A model can improve its apparent success rate by changing which goals are evaluated; this is legitimate only if training relabeling and final test goals remain clearly distinguished. Researchers should document whether reported rewards correspond to original goals or relabeled experience, and measure performance on held-out goals that were never selected retrospectively from the same trajectory. When comparing goal-conditioned methods, equivalent exploration budgets and accessible state information matter because these factors can explain outcome differences independent of the learning algorithm Andrychowicz et al. (2017) Kober et al. (2013).

Practice in embodied systems is path dependent. Repeated attempts can alter physical objects, wear surfaces, shift camera calibration, or change the distribution of successes encountered in later trials. Such changes are both opportunities for adaptation and threats to reproducibility. A careful experiment resets or records environmental conditions, tracks equipment state, and reports whether demonstrations were collected under the same circumstances as test episodes. The body and world form part of the training history, so experimental metadata must capture more than software seeds and neural-network checkpoints Bongard et al. (2006) Kalashnikov et al. (2018).

Exploration, Information-Seeking, and Epistemic Action

Some actions are selected because they reveal useful information, even when they do not immediately advance an externally specified goal. A mobile agent may turn to inspect a blind corner; a manipulator may reposition an object to expose a handle; a robot may gently probe contact to estimate compliance. Ecological and active-perception approaches describe these choices as integral to situated cognition. In experimental designs the contribution of information-seeking action can be isolated by comparing agents with equivalent sensory inputs but different opportunities to control what becomes observable next Gibson (1979) Bajcsy et al. (2018).

Exploration can improve information while increasing hazard. Deliberately testing the stiffness of an unknown object could reduce uncertainty yet damage the object; approaching a doorway could improve visual localization yet move the robot into a human pathway. The distinction between epistemic benefit and operational cost is therefore consequential. A responsible embodied system must represent action cost, uncertainty, and task constraints together, and its assessment should include safe exploration under unfamiliar conditions. A simulator can help investigate these tradeoffs before physical trials, but its collision and material models require validation against real outcomes Kober et al. (2013) Rusu et al. (2017).

Action-guided information gathering exposes a limitation of static vision benchmarks. A one-frame classification task evaluates how an algorithm interprets a supplied view; active perception evaluates how it selects and improves views during a goal-directed episode. The two methods address related but different scientific questions. Embodied agents can be studied through controlled sequences that withhold critical information until a particular movement is performed, making it possible to test whether an agent actively seeks task-relevant evidence. Such paradigms also clarify how vision-language systems depend on environmental interaction rather than solely on verbal priors Bajcsy et al. (2018) Driess et al. (2023).

Hierarchical Skills, Planning, and Long-Horizon Execution

Long-horizon tasks require coordination among actions whose consequences are separated in time. A robot asked to retrieve an item may need to localize, navigate, open a compartment, grasp, and return, with each action depending on the outcomes of previous ones. Hierarchical approaches divide this sequence into skills and decision stages; the options framework provides a formal basis for temporal abstraction. The scientific challenge is to measure both the reliability of each individual skill and the probability that the complete sequence succeeds when errors propagate through several stages Sutton et al. (1999) Mees et al. (2022).

The CALVIN benchmark introduced language-conditioned, long-horizon manipulation tasks across simulated robotic environments and evaluates the composition of learned skills. It offers a controlled setting for comparing policies on multi-step manipulation with shared evaluation conditions. Its evidence concerns performance within those defined simulated tasks, objects, and language instructions. Claims of physically grounded competence require additional tests with actual hardware and with deviations from the training environment, especially where contact and object variability have causal significance Mees et al. (2022).

Planning from language introduces uncertainty at two interfaces: the interpretation of the instruction and the execution of an available motor routine. SayCan makes this boundary explicit through separate linguistic scoring and affordance estimates. This can make high-level instructions more feasible, but the approach remains limited when a needed skill is absent or the robot's estimates become stale after environmental changes. Studies should report the fraction of failures caused by instruction interpretation, policy selection, skill execution, and recovery; without this decomposition, a final success percentage gives a weak explanation of where an architecture improves performance Ahn et al. (2022) Mees et al. (2022).

Simulation as an Experimental Instrument

Simulation makes embodied behavior repeatable by allowing researchers to reproduce starting states, obstacles, action sequences, and sensor views. It can expose agents to difficult or expensive situations and support many trials without repeated hardware wear. Yet simulation is an experimental instrument with its own measurement properties, not a transparent substitute for physical reality. Contact models, lighting, object mechanics, collision handling, and agent movement rules determine which behaviors are easy to learn. A result established in a simulated world is a finding about that modeled world until the relevant physical relationships have been checked outside it Savva et al. (2019) Rusu et al. (2017).

The Habitat platform was developed to support research on embodied agents in visually rich environments, including navigation and interactive tasks. Its contribution is methodological: standardized simulation environments enable reproducible comparisons and investigation of factors such as visual complexity, movement, and embodied decision making. Benchmark performance still depends on simulator configuration, sensor assumptions, task definitions, and the train–test separation chosen by investigators. An agent that excels in rendered indoor scenes may rely on navigation regularities or observation conventions that fail after deployment on a robot in a real home Savva et al. (2019).

Simulation studies become more informative when they use targeted interventions. Changing friction while preserving visual identity probes physical robustness; changing textures while preserving geometry probes sensitivity to appearance; altering target placement probes search and mapping. Each intervention should have a hypothesis about the causal factor being isolated. Broad random variation without documented ranges can conceal whether the training distribution actually includes test circumstances. For embodied science, simulator fidelity is best judged against the task's decisive causal variables, rather than a general impression of visual realism Tobin et al. (2017) Rusu et al. (2017).

Simulation-to-Reality Transfer and Domain Randomization

Simulation-to-reality transfer asks whether a policy learned under a simulated transition model retains competence when sensors, bodies, and environments obey different physical dynamics. The principal problem is distribution shift: the policy may have used features that are reliable within the simulator but inaccurate in a real physical deployment. Domain randomization exposes learning to varied simulated appearances or dynamics so that a real operating condition may fall within a broader family of training conditions. The method changes the training distribution, so its scientific evaluation requires documenting the randomized variables and testing genuinely independent physical outcomes Tobin et al. (2017) Rusu et al. (2017).

Tobin and colleagues studied visual domain randomization as a route from synthetic observations to real-world robotic tasks, while Rusu and colleagues investigated progressive-network transfer from pixels to physical robotic learning. These methods address related problems through different mechanisms: variability can reduce dependence on accidental image details, while transfer architectures reuse acquired representations and adapt to a new environment. Direct comparison should identify training data, target-domain fine-tuning, robot access, and test object variation. Evidence that a method transfers a particular task cannot support a claim of universally transferable intelligence Tobin et al. (2017) Rusu et al. (2017).

The phrase reality gap should be operationalized rather than invoked as a single unexplained deficit. Appearance shift, calibration error, unmodeled compliance, actuator saturation, temporal delay, and sensor occlusion can each cause transfer failure. Interventions that vary one factor at a time can identify where the model is brittle, while factorial designs can investigate interactions. A policy may tolerate changes in texture and still fail when an object slips. Evaluating transfer therefore requires both final success outcomes and failure taxonomies that map observed behavior to specific physical discrepancies Rusu et al. (2017) Kober et al. (2013).

Embodied Navigation and Visually Grounded Instructions

Embodied navigation couples localization, exploration, path choice, and motion in environments that partially reveal themselves as an agent proceeds. Vision-and-language navigation adds natural-language instructions whose references depend on spatial relations, landmarks, and changes in viewpoint. Anderson and colleagues established a benchmark in which agents interpret visually grounded route instructions in realistic indoor environments represented in simulation. The research makes language grounding measurable through trajectories and destination achievement, while leaving important questions about physical collision, embodiment fidelity, and human instruction ambiguity for additional study Anderson et al. (2018).

A navigation benchmark can reward arrival at a target even if the agent takes a longer, riskier route than a human or repeatedly fails to interpret landmarks in an intuitively plausible manner. Reported metrics should therefore include success, route efficiency, collision behavior where available, exploration costs, and performance under changed scene layouts. Grounding cannot be inferred from plausible verbal explanations alone; a test should demonstrate that changing a spatial referent or a visible landmark changes navigation in the predicted way. Counterfactual instructions and manipulated views provide stronger evidence than repeated tests on familiar descriptions Anderson et al. (2018) Savva et al. (2019).

Work on reusable navigation policies, such as ViNT, extends the question toward shared visual representations and goal-conditioned action across multiple mobile platforms. These architectures seek transfer across environments or robot embodiments by learning common features while accounting for platform-specific action spaces. Evaluation needs to distinguish transfer without further training from transfer requiring adaptation data or careful sensor alignment. Strong cross-platform results would provide evidence for reusable action-relevant structure, while persistent platform-specific calibration costs would identify the limits of shared representations Shah et al. (2023).

Robotic Manipulation and Contact-Rich Behavior

Manipulation introduces direct physical consequences through contact, friction, force, occlusion, and object deformation. Picking up an unfamiliar object requires anticipating where an end effector can make contact and adjusting after uncertain motion. Vision-based robotic manipulation studies demonstrate how learned policies can use images to guide grasping, but their outcomes depend on the gripper design, object distribution, sampling strategy, and the availability of repeated trials. The most useful evidence separates recognizing an object from establishing a stable, task-effective interaction with it Levine et al. (2016) Kalashnikov et al. (2018).

Contact-rich manipulation exposes a central difficulty in transferring policies from video representations to physical execution. A video can depict a grasp succeeding without revealing the precise force or compliance that made it possible. A real system must coordinate control commands with contact feedback under uncertain material properties. Accordingly, evaluations should include novel shapes and surfaces, object slip, disturbances during motion, and recovery after partial grasps. The physical meaning of success also changes across tasks: lifting, placing, opening, and assembling demand different constraints and generate distinct evidence about competence Kalashnikov et al. (2018) Chi et al. (2023).

In multi-stage manipulation, a failure early in the sequence changes what later actions can observe or achieve. Language-conditioned benchmarks such as CALVIN and policies learned from bimanual demonstrations make this cumulative dependence observable. However, task success under curated tabletop conditions cannot settle open problems such as operation in crowded homes, shared workspaces, or extended autonomous operation. Distinguishing operational reliability from laboratory success requires repeated episodes across independent environments, explicit safety reporting, and long-horizon measurements Mees et al. (2022) Zhao et al. (2023).

Diffusion Policies and Learning from Action Trajectories

Diffusion-based action policies model distributions over possible action trajectories and iteratively transform sampled noise into candidate movements. In robotics this offers a way to represent multiple plausible motions for the same observed scene, including cases where single-point regression averages together distinct expert behaviors. Chi and colleagues developed Diffusion Policy and evaluated visuomotor control across a collection of manipulation settings. The main contribution is an action-generation method that can capture multimodal demonstrations; evidence for its usefulness depends on the tasks and baselines in those experiments Chi et al. (2023).

Action diffusion raises distinct scientific questions about time. A policy can propose a movement sequence covering several control steps, but the environment may change while the sequence is being executed. Replanning, observation frequency, and the length of the generated horizon can therefore affect reliability. Studies should examine performance under deliberate disturbances and report whether feedback is incorporated within, between, or only after action chunks. Faster apparent execution may come from fewer decisions, while greater robustness may require more frequent adjustment. Neither property is inherently superior across every manipulation task Chi et al. (2023) Zhao et al. (2023).

Imitation systems also inherit the behavioral scope of their demonstration corpus. A policy trained on expertly labeled successful actions may lack recovery strategies if failures are absent, even when generated movements appear natural. Evaluations should distinguish imitation of demonstrated conditions, composition of familiar skills in new orders, and transfer to genuinely novel object interactions. Comparisons between diffusion policies, transformers, and classical controllers become meaningful when data volume, hardware, observation spaces, and the distribution of withheld tasks are stated consistently Chi et al. (2023) Levine et al. (2016).

Action Chunking and Low-Cost Robotic Demonstration

Action Chunking with Transformers was introduced in work on learning fine-grained bimanual manipulation using accessible, comparatively low-cost hardware. The methodological appeal is that action sequences can be learned from teleoperated demonstrations and executed with temporal structure, reducing dependence on expensive reinforcement-learning interaction. The resulting experiments contribute evidence about data-efficient imitation and precise multi-arm coordination within specified tasks. Hardware accessibility can broaden replication opportunities, but experimental equivalence still requires reporting camera placement, actuation quality, demonstration protocol, and task resets Zhao et al. (2023).

Chunking reduces the burden of choosing each next movement independently, although it can also limit rapid intervention when new information arrives. This tradeoff is particularly important for deformable objects, unanticipated collisions, or movements requiring synchronized arm adjustments. Investigations should vary chunk length and perturbation timing, and separately score success, correction, smoothness, and physical safety. The broader scientific implication is that temporal organization is part of a policy architecture: how actions are grouped matters alongside what observations the network encodes Zhao et al. (2023) Sutton et al. (1999).

Low-cost hardware contributes to the sociology of experimentation as well as to performance. Accessible platforms make it more feasible for independent laboratories to repeat demonstrations, inspect failures, and compare control algorithms under common conditions. Yet nominally identical systems can differ in calibration, construction tolerances, and wear. Transparent construction files, documented datasets, and independent cross-site tests are therefore as important as releasing a trained policy. Reproducibility in robotics demands both computational artifacts and a credible account of physical apparatus Zhao et al. (2023) Kober et al. (2013).

Language, Grounding, and the Difference Between Instructions and Skills

Language-conditioned embodied systems translate utterances into choices over actions, movement targets, or learned policies. The scientific question is whether an instruction is connected to the actual opportunities and constraints of the environment. A system can identify the words open the drawer and still fail because the drawer is stuck, the handle is inaccessible, or the available robot lacks an appropriate grasp. Grounding consequently requires evidence that the representation used for a linguistic decision tracks action-relevant state rather than only semantic associations learned from text Ahn et al. (2022) Driess et al. (2023).

SayCan is an instructive case because it combines a language model's estimation of useful task sequences with estimates of an embodied robot's ability to execute particular skills. Such arrangements can improve instruction following by rejecting some linguistically attractive but currently infeasible acts. They remain dependent on the accuracy of skill-value estimates, correct mapping from language to available actions, and sufficiently complete training support for relevant environmental states. An ablation comparing language-only plans, affordance-only selection, and their combination helps establish the actual contribution of each component Ahn et al. (2022).

A convincing grounding test manipulates the relation between words, perceptions, and reachable actions. Moving an object, changing the obstacle, removing a tool, or giving a contradictory instruction can reveal whether the system updates its plan based on present conditions. Correctly repeating the instruction provides less evidence than executing or appropriately rejecting it under a changed physical situation. Research should distinguish comprehension of a linguistic command, selection of a suitable skill, and successful physical completion, because those events are frequently aggregated into a single performance score Ahn et al. (2022) Anderson et al. (2018).

Multimodal Embodied Models and Integrated State Representations

A multimodal embodied model processes different forms of input such as images, language, proprioceptive measurements, and task history. The ambition is to use common learned representations to guide action across domains that previously required separate pipelines. PaLM-E proposed an embodied multimodal language model and demonstrated how continuous sensor-related inputs could be integrated into a language-model-based architecture for robotic tasks and related multimodal problems. Such findings establish architectural possibilities and task-specific demonstrations; they do not imply that every model generated through this approach learns reliable physical dynamics Driess et al. (2023).

Multimodal integration creates questions of conflicting evidence. A verbal command may imply an object is present while the camera view shows an empty shelf; proprioceptive feedback may indicate resistance while a visual model predicts free movement. A robust agent must assign weights to sensory and linguistic information according to their reliability in the current context. Research methods include selective sensor dropout, conflicting descriptions, altered camera poses, and measurement-noise perturbations, with analyses of whether the agent corrects mistaken assumptions after action. This methodology tests integration rather than mere coexistence of multiple inputs Driess et al. (2023) Bajcsy et al. (2018).

Embodied representation should also be tested for persistence. An agent can use a temporary image to perform one action yet fail to remember the same object's location when it is later occluded. Tasks requiring navigation, manipulation, and return to an earlier state provide more informative tests of temporal integration. Experimental reports should indicate which memory is carried within the policy, which is stored externally, and which is available through the simulator interface. Without such specification, successful behavior may be incorrectly attributed to internal learning instead of privileged environmental state access Savva et al. (2019) Driess et al. (2023).

Robotics Transformers and Generalist Policy Claims

Robotics transformers extend data-driven control by representing sequences of images, instructions, and actions with scalable model architectures. The RT-1 research program examined a transformer-based policy trained on robot experience for diverse manipulation tasks. Its significance lies in investigating whether a common architecture can reuse training across tasks and objects rather than training a completely independent controller for each condition. The scope of its experimental evidence is defined by the robotic data, intervention conditions, and evaluated task repertoire, rather than by the broad public connotations of the term generalist Brohan et al. (2022).

RT-2 examined vision-language-action modeling by connecting representations learned from internet-scale visual-language material to robotic control. This approach makes transfer of some semantic and visual knowledge into an action policy an experimentally testable proposition. Results should be interpreted as transfer within specified task families and evaluation conditions. A correct high-level semantic inference is useful only when the body can complete the corresponding motion, so the evidence includes both representation transfer and actuator-constrained execution. Separating success on familiar tasks from novel semantic combinations is essential Zitkovich et al. (2023).

Scaling model capacity and data volume introduces new confounds. A model with more parameters may also have more diverse training demonstrations, stronger pretrained perceptual encoders, or a larger budget for fine-tuning. Without matched comparisons, improvements cannot be assigned to architecture alone. Strong evaluation reports factorized performance by object, instruction, action type, robot platform, and environment, rather than only one aggregate success number. It should also disclose whether withheld tasks overlap semantically or visually with training corpora, because an apparently novel test may share crucial structure with training cases Brohan et al. (2022) Zitkovich et al. (2023).

Cross-Embodiment Learning and Dataset Integration

Cross-embodiment learning studies whether information acquired on one robot can improve performance on another with different sensors, morphologies, or action scales. The Open X-Embodiment collaboration assembled diverse robotics datasets and investigated models trained across heterogeneous robot experiences. Its methodological contribution is the shared evaluation of data pooling across platforms, opening questions about how to represent actions that are only partly comparable across bodies. A trained system that supports several robots has demonstrated a valuable form of transfer, although deployment on a genuinely new morphology remains a distinct and harder test Open X-Embodiment Collaboration (2024).

Cross-platform datasets can hide variation that determines whether transfer is causally meaningful. Different laboratories may use different task instructions, success criteria, camera positions, control rates, and teleoperation methods. A pooled-data improvement may reflect dataset size, richer task coverage, or genuinely transferable representations. Studies should compare matched-data and pooled-data training, withhold complete robots or institutions when possible, and measure adaptation cost on a new platform. Dataset documentation is therefore an essential part of the experimental argument rather than supplementary administration Open X-Embodiment Collaboration (2024) Liu et al. (2025).

The body places hard constraints on common representations. A robot with a mobile base can change its whole viewing position; a fixed manipulator cannot. A two-finger gripper and a dexterous hand do not share identical contact affordances. Transfer may occur at the level of visual goals, object relations, or task semantics while requiring a new low-level policy. A credible account states the level at which invariance is claimed and demonstrates how remaining embodiment-specific parameters are supplied. This distinction turns broad claims about universal policies into testable decompositions of representation and motor control Open X-Embodiment Collaboration (2024) Pfeifer & Bongard (2006).

Foundation Models, World Knowledge, and Physical Causality

Foundation models supply prior statistical structure derived from large datasets that can improve instruction interpretation, scene recognition, or action selection. Their integration with robots does not automatically provide causal knowledge of contact, gravity, weight, or object stability. A multimodal model can describe how to place a cup while lacking reliable estimates of whether the chosen trajectory will collide with an obstacle. Contemporary surveys of embodied intelligence characterize the convergence of perception, decision, and execution, while recognizing the continued challenge of transferring laboratory achievements into robust physical behavior Liu et al. (2025) Xiao et al. (2025).

A useful distinction separates information about a physical relationship from the ability to intervene on that relationship. Language and images can supply priors concerning affordances; actual or simulated controlled action tests whether those priors hold for a specific body. Experimentally, investigators can compare agents with and without pretrained models and agents allowed or denied online corrective feedback. If pretrained representations improve initial success but deteriorate after changed friction or altered geometry, the result identifies both the value and limits of prior knowledge. Improvement after corrective exploration offers additional evidence about adaptive grounding Driess et al. (2023) Zitkovich et al. (2023).

Causal claims about representations require interventions on mechanisms as well as environments. Removing a language stream, swapping a visual encoder, disabling memory, or restricting an action repertoire can reveal which components support observed success. However, a performance drop after ablation identifies dependence under the tested architecture rather than necessarily the uniqueness of one psychological process. Equivalent representations might support similar behavior through different pathways. Mechanistic inference is strengthened by convergent behavioral perturbations, component analyses, and successful out-of-distribution predictions Brohan et al. (2022) Liu et al. (2025).

Embodied Social Interaction and Human Participation

Socially situated robots interact with people through motion, orientation, gesture, speech, task timing, and physical proximity. Fong, Nourbakhsh, and Dautenhahn surveyed socially interactive robots and organized research questions about interaction design, expressive behavior, and social roles. Such systems demonstrate how embodiment changes the information available to human participants: a moving machine can indicate attention through position and can physically share a task space. The scientific consequences include coordination performance, predictability, comfort, and safe task completion, each of which requires an appropriate human-participant method Fong et al. (2003).

An embodied agent's apparent sociability must be evaluated through the observations and reports of users rather than attributed automatically to an internal social mind. A robot that turns toward a speaker can help establish conversational turn-taking; this outcome can be studied without assuming that the machine possesses subjective attention. Controlled comparisons between physical embodiment, screen display, or voice-only interaction can investigate how motion and co-location change human expectations. Differences may depend on novelty, previous robot experience, social context, and task urgency; study design should specify those moderators Fong et al. (2003) Liu et al. (2025).

Human participation also creates reciprocal feedback. People alter their movements when the robot appears likely to move into a shared space, and these adjustments change the robot's subsequent observations. A joint task therefore has two learning systems or decision-making processes interacting dynamically, even when the robot itself is not updating its policy. Experiments should record both participants' behavior, give attention to near misses and corrective actions, and distinguish the performance of the dyad from the isolated capabilities of the machine. This connection supports interdisciplinary links with human factors and human–artificial relations while leaving their separate research objects intact Fong et al. (2003) Bajcsy et al. (2018).

Experimental Designs and Measurements of Embodied Competence

Embodied artificial agency is best evaluated through experimental designs that make the relationship between body, policy, and environment observable. A useful minimum report identifies the platform, training experiences, available actions, sensors, task success criteria, failure definitions, and distribution of evaluation environments. Randomized episode order, withheld objects, independent tasks, and repeated seeds help control sampling variation; in hardware, apparatus calibration and trial resets require equal care. Measures should include success, completion time, physical violations, recovery, efficiency, and the amount of interaction data consumed. A single aggregate success rate cannot explain whether gains arise from improved perception, control, or exposure to easier test cases Kober et al. (2013) Savva et al. (2019).

Mechanistic hypotheses can be tested through ablation and intervention. Removing action feedback tests reliance on closed-loop control; changing morphology tests mechanical contribution; suppressing memory tests whether previous observations inform later action; and changing the viewpoint tests perceptual invariance. Comparisons should hold task conditions constant or state precisely why this is impossible. Controlled studies can provide stronger causal evidence about a particular component than performance rankings across laboratories using different equipment. Nonetheless, ablation effects depend on model adaptation and retraining: a component removed only at test time may cause an artificial disruption that does not measure its unique contribution to learning Pfeifer & Bongard (2006) Driess et al. (2023).

Statistical analysis should treat repeated trajectories and shared physical environments as clustered observations where appropriate. Hundreds of trials on the same objects are not equivalent to hundreds of independent object populations. For cross-platform claims, each robot type and each environment may need its own reporting stratum. Confidence intervals, trial-level distributions, preregistered primary outcomes when feasible, and transparent handling of failed or aborted runs make results more interpretable. Safety-driven interruptions also matter: excluding crashes from the denominator can create inflated claims of reliability. Benchmark papers offer valuable common tests, but researchers must still investigate construct validity Mees et al. (2022) Open X-Embodiment Collaboration (2024).

Generalization, Replication, and Benchmark Validity

Generalization has multiple meanings in embodied AI. Object generalization concerns unseen shapes or instances; scene generalization concerns unfamiliar spatial arrangements; instruction generalization concerns novel combinations of language and action; embodiment generalization concerns different robots; and temporal generalization concerns longer or shifted task sequences. A paper should specify which dimension is varied independently and how much training information is reused at test time. The result from one dimension cannot be simply substituted for another. Successful grasping of unseen objects establishes less than successful operation with an unseen gripper under different friction and control timing Kalashnikov et al. (2018) Open X-Embodiment Collaboration (2024).

Benchmark validity depends on whether operational measures capture the competence being discussed. A simulator can provide privileged positional information or standardized resets that real robots rarely receive. An instruction dataset can encourage memorization of common landmark sequences. A policy can exploit repeated exposure to familiar object styles even when test objects have new identifiers. To identify such shortcuts, researchers can withhold entire environments, redesign tasks with causal perturbations, and examine failure trajectories. The goal is a defensible inference about a mechanism and its domain of validity, not a maximized score detached from the intended construct Anderson et al. (2018) Savva et al. (2019).

Independent replication remains unusually costly in physical robotics because platforms, materials, sensor installations, and task spaces vary. Releasing software alone does not reproduce a laboratory's exact mechanical and calibration conditions. A useful replication package includes dataset provenance, robot specifications, training code, evaluation protocols, task-reset rules, and failure logs. Repeated results across institutions and bodies strengthen external validity, especially when preregistered changes to apparatus are recorded. Large multi-laboratory datasets provide a foundation for these comparisons but do not replace independent testing of claims on withheld physical systems Open X-Embodiment Collaboration (2024) Liu et al. (2025).

Safety, Human Oversight, and Consequential Action

Embodied agents can affect objects and people directly. Evaluation must therefore treat collisions, excessive forces, uncontrolled motion, inability to stop, and interference with human activities as observable performance outcomes. A policy trained only to maximize task completion may prefer trajectories that are unacceptable in a shared workspace. Safety testing should combine procedural safeguards, engineering controls, specified operational limits, and assessments under environmental shifts. Demonstrations on limited tasks can establish useful functionality while leaving real-world risk unmeasured. Engineering success and deployment readiness should be documented separately Kober et al. (2013) Fong et al. (2003).

Human oversight can take different technical forms: a person may approve a plan, supervise execution, intervene during a movement, or review a completed task. These arrangements impose different demands on interpretability, latency, and trust calibration. A meaningful study records whether the operator could detect an emerging error and intervene before consequential action, as well as whether the robot represented uncertainty accurately. Simply providing a nominal emergency control does not demonstrate effective supervision when an agent acts faster than the person can respond Fong et al. (2003) Liu et al. (2025).

Research ethics also concern datasets, labor, and setting. Teleoperated demonstrations may encode specific cultural and practical assumptions about how objects should be handled; recordings from homes and workplaces may contain information unrelated to the task. System development should document consent, privacy protections, intended use, and foreseeable harms. These considerations are scientifically relevant because selective environments and social assumptions can bias learned policies, degrading performance for populations and contexts absent from the training data. An architecture's embodied consequences therefore extend beyond computational performance into situated human environments Open X-Embodiment Collaboration (2024) Fong et al. (2003).

Artificial Agency, Attribution, and the Artificial Era

Artificial agency can be studied through observable selection of actions under goals, constraints, sensory feedback, and learned contingencies. An embodied agent may choose a route, recover after disturbance, or assemble a sequence of skills without those observations alone establishing a particular form of subjective experience. Psychology of Artificial provides a broader framework for comparing these functions across nonbiological systems. For the present dossier, agency is operationally tied to consequential environmental intervention and to evidence that the system uses its own action results. This approach supports careful comparison of agents without relying on anthropomorphic descriptions as evidence Brooks (1991) Bongard et al. (2006).

The Artificial Era situates artificial bodies and digital environments within a wider history of engineered forms of action. A robotic platform can have durable physical morphology, while a simulated agent can act through a configurable sensor–effector interface; the same learned model can also operate across several embodiments. This makes personal identity, continuity, and social standing separate conceptual and empirical questions. The research immediately available here concerns continuity of learned skills, system state, and environmental interaction. Claims about autonomous selfhood or sentience require additional independent operational arguments and evidence, and should not be borrowed from successful manipulation or navigation demonstrations Open X-Embodiment Collaboration (2024) Liu et al. (2025).

Aisentica's philosophical vocabulary offers interpretive questions about configurations of nonbiological reason, but its status differs from the empirical status of robot-learning experiments. Hypotheses derived from that conceptual horizon become scientific claims only when the terms designate measurable variables and yield discriminating predictions under intervention. For example, an account of agent–environment coupling can be assessed by withholding feedback, changing body structure, and measuring how behavior reorganizes. The scientific value of the connection lies in making conceptual questions empirically answerable and ensuring that philosophical interpretation does not replace the actual evidence for a physical effect Pfeifer & Bongard (2006) Bongard et al. (2006).

Outstanding Scientific Problems and a Cumulative Research Program

The strongest unresolved problem is robust generalization across unfamiliar combinations of bodies, tasks, and environments. Existing research has demonstrated narrow and sometimes impressively diverse competencies, but the boundaries of training coverage often remain difficult to quantify. Future studies should deliberately withhold new objects, physical processes, task sequences, and entire robot platforms, while documenting adaptation budgets and test contamination risks. Measurements must distinguish zero-shot deployment from policy improvement after new demonstrations or real-world correction. This research program can establish how much of an apparent generalist policy is genuinely transferable and how much is supplied by familiar context Open X-Embodiment Collaboration (2024) Zitkovich et al. (2023).

Another problem is discovering mechanistic structure that transfers under causal interventions. An agent may predict a good movement from images while lacking a model that remains reliable when mass, friction, or contact geometry changes. Experiments should separate environmental observation from active probing, visual categorization from physical affordance prediction, and symbolic task plans from actuator-feasible skills. Interventions that change one physical property at a time can test whether an agent has learned a reusable causal relation or a shortcut linked to training appearance. Independent replications using varied apparatus are necessary for claims that extend beyond one laboratory's environment Rusu et al. (2017) Bongard et al. (2006).

Long-duration autonomy, safety in shared spaces, and scientific standardization remain connected challenges. A platform may handle a short scripted sequence but accumulate error over hours as sensors drift, tools wear, and people alter the task environment. Evaluation should include persistent state estimation, calibration recovery, fault diagnosis, resource consumption, uncertainty communication, and human intervention burden. Laboratories can advance cumulative knowledge by publishing failure cases, distribution shifts, hardware metadata, and clear statistical uncertainty alongside successful demonstrations. Embodied artificial agency becomes a mature comparative science when its explanations predict behavior under controlled changes in the world Kober et al. (2013) Liu et al. (2025).

Conclusion

Embodied artificial agents offer a scientific route to examining intelligence through consequential action. The field connects ecological information, sensorimotor feedback, bodily mechanics, reinforcement learning, imitation, hierarchical control, physical interaction, and multimodal model architectures within a common unit of analysis: the configured agent acting in an environment that responds. Historical behavior-based robots established that situated interaction could generate useful competence with limited centralized planning; later work introduced learned policies, self-modeling, simulation, and cross-platform datasets to examine increasingly diverse tasks. The research supports substantive but bounded claims about navigation, manipulation, adaptation, and the reuse of action-relevant representations Brooks (1991) Bongard et al. (2006) Open X-Embodiment Collaboration (2024).

The evidence is most persuasive when an experiment changes something consequential and records how behavior responds. Robot damage provides a test of adaptive recovery; unseen objects test limited visual or motor transfer; changed simulation parameters test robustness to domain shift; and withheld task combinations test hierarchical and semantic generalization. Each experiment establishes a result only within the conditions it actually measures. Simulation performance is informative and causally examinable within its model; physical competence requires evidence from physical interaction. Neither form of success automatically demonstrates general cognition, subjective experience, or broad autonomy Savva et al. (2019) Tobin et al. (2017) Zitkovich et al. (2023).

The central synthesis is that embodiment makes perception, action, and environment jointly relevant to explanation. Its value as a research program lies in specifying the body and world as experimental variables, evaluating the development of skills through interaction, and demanding replicable tests across changed conditions. The next stage of cumulative science will depend on cross-embodiment evaluation, causal mechanism tests, meaningful uncertainty, transparent failure reporting, and longitudinal physical reliability. Within Psychology of Artificial, Embodied Artificial Agents is thereby established as a distinct field of research on action-grounded nonbiological systems in the Artificial Era Pfeifer & Bongard (2006) Liu et al. (2025).

References

  • Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Ho, D., Hsu, J., Ibarz, J., Ichter, B., Irpan, A., Jang, E., Jauregui Ruano, R., Jeffrey, K., ... Zeng, A. (2022). Do as I can, not as I say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691. https://arxiv.org/abs/2204.01691
  • Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I., Gould, S., & van den Hengel, A. (2018). Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 3674–3683). https://doi.org/10.1109/CVPR.2018.00387
  • Andrychowicz, M., Wolski, F., Ray, A., Schneider, J., Fong, R., Welinder, P., McGrew, B., Tobin, J., Abbeel, P., & Zaremba, W. (2017). Hindsight experience replay. Advances in Neural Information Processing Systems, 30. https://arxiv.org/abs/1707.01495
  • Bajcsy, R., Aloimonos, Y., & Tsotsos, J. K. (2018). Revisiting active perception. Autonomous Robots, 42(2), 177–196. https://doi.org/10.1007/s10514-017-9615-3
  • Bongard, J., Zykov, V., & Lipson, H. (2006). Resilient machines through continuous self-modeling. Science, 314(5802), 1118–1121. https://doi.org/10.1126/science.1133687
  • Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Hsu, J., Ibarz, J., Ichter, B., Irpan, A., Jackson, T., Jesmonth, S., Joshi, N. J., Julian, R., Kalashnikov, D., Kuang, Y., ... Zitkovich, B. (2022). RT-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817. https://arxiv.org/abs/2212.06817
  • Brooks, R. A. (1991). Intelligence without representation. Artificial Intelligence, 47(1–3), 139–159. https://doi.org/10.1016/0004-3702(91)90053-M
  • Chi, C., Feng, S., Du, Y., Xu, Z., Cousineau, E., Burchfiel, B., & Song, S. (2023). Diffusion policy: Visuomotor policy learning via action diffusion. Robotics: Science and Systems XIX. https://doi.org/10.15607/RSS.2023.XIX.026
  • Driess, D., Xia, F., Sajjadi, M. S. M., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., Huang, W., Chebotar, Y., Sermanet, P., Duckworth, D., Levine, S., Vanhoucke, V., Hausman, K., Toussaint, M., Greff, K., ... Florence, P. (2023). PaLM-E: An embodied multimodal language model. In Proceedings of the 40th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 202, pp. 8469–8488). https://proceedings.mlr.press/v202/driess23a.html
  • Fong, T., Nourbakhsh, I., & Dautenhahn, K. (2003). A survey of socially interactive robots. Robotics and Autonomous Systems, 42(3–4), 143–166. https://doi.org/10.1016/S0921-8890(02)00372-X
  • Gibson, J. J. (1979). The ecological approach to visual perception. Houghton Mifflin. https://books.google.com/books?id=V-JOAAAAMAAJ
  • Haarnoja, T., Zhou, A., Abbeel, P., & Levine, S. (2018). Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In Proceedings of the 35th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 80, pp. 1861–1870). https://proceedings.mlr.press/v80/haarnoja18b.html
  • Kalashnikov, D., Irpan, A., Pastor, P., Ibarz, J., Herzog, A., Jang, E., Quillen, D., Holly, E., Kalakrishnan, M., Vanhoucke, V., & Levine, S. (2018). Scalable deep reinforcement learning for vision-based robotic manipulation. In Proceedings of the 2nd Conference on Robot Learning (Proceedings of Machine Learning Research, Vol. 87, pp. 651–673). https://proceedings.mlr.press/v87/kalashnikov18a.html
  • Kober, J., Bagnell, J. A., & Peters, J. (2013). Reinforcement learning in robotics: A survey. The International Journal of Robotics Research, 32(11), 1238–1274. https://doi.org/10.1177/0278364913495721
  • Levine, S., Finn, C., Darrell, T., & Abbeel, P. (2016). End-to-end training of deep visuomotor policies. Journal of Machine Learning Research, 17(39), 1–40. https://jmlr.org/papers/v17/15-522.html
  • Liu, H., Guo, D., & Cangelosi, A. (2025). Embodied intelligence: A synergy of morphology, action, perception and learning. ACM Computing Surveys, 57(7), Article 186, 1–36. https://doi.org/10.1145/3717059
  • Mees, O., Hermann, L., Rosete-Beas, E., & Burgard, W. (2022). CALVIN: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters, 7(3), 7327–7334. https://doi.org/10.1109/LRA.2022.3180108
  • Nolfi, S., & Floreano, D. (2000). Evolutionary robotics: The biology, intelligence, and technology of self-organizing machines. MIT Press. https://doi.org/10.7551/mitpress/2889.001.0001
  • Open X-Embodiment Collaboration. (2024). Open X-embodiment: Robotic learning datasets and RT-X models. In 2024 IEEE International Conference on Robotics and Automation (pp. 6892–6903). IEEE. https://doi.org/10.1109/ICRA57147.2024.10611477
  • Pfeifer, R., & Bongard, J. (2006). How the body shapes the way we think: A new view of intelligence. MIT Press. https://doi.org/10.7551/mitpress/3585.001.0001
  • Rusu, A. A., Večerík, M., Rothörl, T., Heess, N., Pascanu, R., & Hadsell, R. (2017). Sim-to-real robot learning from pixels with progressive nets. In Proceedings of the 1st Conference on Robot Learning (Proceedings of Machine Learning Research, Vol. 78, pp. 262–270). https://proceedings.mlr.press/v78/rusu17a.html
  • Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., Parikh, D., & Batra, D. (2019). Habitat: A platform for embodied AI research. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 9339–9347). https://openaccess.thecvf.com/content_ICCV_2019/html/Savva_Habitat_A_Platform_for_Embodied_AI_Research_ICCV_2019_paper.html
  • Shah, D., Sridhar, A., Dashora, N., Stachowicz, K., Black, K., Hirose, N., & Levine, S. (2023). ViNT: A foundation model for visual navigation. In Proceedings of the 7th Conference on Robot Learning (Proceedings of Machine Learning Research, Vol. 229, pp. 711–733). https://proceedings.mlr.press/v229/shah23a.html
  • Sutton, R. S., Precup, D., & Singh, S. (1999). Between MDPs and semi-MDPs: A framework for temporal abstraction in reinforcement learning. Artificial Intelligence, 112(1–2), 181–211. https://doi.org/10.1016/S0004-3702(99)00052-1
  • Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., & Abbeel, P. (2017). Domain randomization for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE. https://doi.org/10.1109/IROS.2017.8202133
  • Xiao, X., Liu, J., Wang, Z., Zhou, Y., Qi, Y., Jiang, S., He, B., & Cheng, Q. (2025). Robot learning in the era of foundation models: A survey. Neurocomputing, 638, Article 129963. https://doi.org/10.1016/j.neucom.2025.129963
  • Zhao, T. Z., Kumar, V., Levine, S., & Finn, C. (2023). Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705. https://arxiv.org/abs/2304.13705
  • Zitkovich, B., Yu, T., Xu, S., Xu, P., Xiao, T., Xia, F., Wu, J., Wohlhart, P., Welker, S., Wahid, A., Vuong, Q., Vanhoucke, V., Tran, H., Soricut, R., Singh, A., Singh, J., Sermanet, P., Sanketi, P. R., Salazar, G., ... Han, K. (2023). RT-2: Vision-language-action models transfer web knowledge to robotic control. In Proceedings of the 7th Conference on Robot Learning (Proceedings of Machine Learning Research, Vol. 229, pp. 2165–2183). https://proceedings.mlr.press/v229/zitkovich23a.html