Human-Machine Communication

• Human-Machine Communication is a framework where humans and machines share intent, state, and control through iterative loops across conversational, embodied, and networked modalities.
• It integrates adaptive interfaces, shared control, and AI-in-the-loop mechanisms which enhance safety, performance, and decision-making in complex systems.
• Empirical evaluations employ metrics like prediction accuracy and coherence measures to optimize design in wireless, adaptive, and embedded HMC setups.

Human-Machine Communication (HMC) denotes the processes through which humans and machines exchange intent, state, meaning, and control across conversational, embodied, and networked settings. In one influential formulation, interaction is a recurring loop that may include human intent formation, expression of that intent, AI inference and clarification, presentation of inferred meaning or execution, and human recognition of alignment or misalignment, with repetition until the human is satisfied (Glassman, 2023). In engineered collaborative systems, HMC also includes runtime arrangements in which AI remains the primary “working horse” but calls on human input when confidence is low or situations become ambiguous, risky, or safety-critical, as well as shared-control regimes in which haptic, visual, neural, and contextual signals are exchanged to coordinate action (Schöning et al., 2023, Lv et al., 2020).

1. Conceptual foundations and control paradigms

A central theme in contemporary HMC is that communication is not reducible to command issuance. The conversational framework for human-computer communication decomposes interaction into a loop comprising: forming an intent, expressing that intent as a command or utterance, AI inference and ambiguity resolution, AI presentation of inferred meaning and/or execution on current and future situations or data, and human recognition of whether the interpretation aligns with intent; along the way, humans may update their model of the AI’s capabilities, the operating situation, their own intent, and the expression they use to communicate it (Glassman, 2023). This formulation places communicative failure not only at the point of decoding, but also at intent formation, feedback perception, and mental-model revision.

In embedded and edge AI, a related control paradigm is “AI-in-the-loop,” which differs from conventional human-in-the-loop practice. The paper defining this approach argues that HMI is usually absent during AI architecture design and model training even though humans already participate in data analysis, data selection and cleaning, and performance evaluation. Its proposed alternative places HMI both before deployment and during inference: during design, humans examine architectural productivity before expensive training; during operation, the AI solves the task autonomously but requests human input when its inference is unsure or when safety or security concerns arise (Schöning et al., 2023). This suggests a shift from static automation levels toward conditional collaboration, where communication is invoked selectively rather than continuously.

Task-allocation research in Industry 4.0 introduces a further distinction between “skill-based intuitive behavior” and “knowledge-based intellectual behavior,” described as two typical modes of human behavior in HMI. On that basis, “intuitive interaction flow” is defined as “a highly integrated and seamless working state formed through dynamic interaction between human intuitive behavior and machine intelligence.” In the corresponding dual-loop model, the intuitive loop emphasizes direct perception-action coupling, procedural memory, minimal cognitive load, and minimal machine assistance, whereas the intellectual loop emphasizes conscious decision-making, explicit rules, higher cognitive load, and greater machine support (Xu et al., 2024). A plausible implication is that HMC design must allocate communicative burden differently depending on whether the operator is acting in a practiced, embodied regime or an analytic, problem-solving regime.

2. Interface design, adaptation, and explanation

A recurring engineering objective in HMC is to expose internal machine state in a form that changes operator behavior before downstream costs accumulate. In AI architecture design, Receptive Field Analysis (RFA) is used to visualize CNN productivity through graph-based visualizations that mark layers in red when they are unproductive and in orange when productivity is borderline. The RFA-Toolbox provides this analysis before training begins, allowing users to redesign architectures such as ResNet18 variants in light of input resolution and task requirements. In the reported user study, without the UI, 68.8% of subjects believed deeper models were more accurate, whereas with the RFA-based HMI, 50% correctly identified the leaner, better-performing architecture; Appendix A also reports feasible input resolutions for architectures such as ResNet18 from 139×139 to 435×435 (Schöning et al., 2023). The immediate significance is not merely usability, but resource-aware model selection in embedded systems.

Adaptive interface research generalizes this logic from model design to runtime guidance. One ongoing approach defines the transformation of user interactions into adaptive HMIs as a data-driven process that logs interface events and contextual variables such as user role, shift, or machine state, extracts ordered sequences $E = \langle e_1, e_2, \ldots, e_m \rangle$, and predicts the next likely action with first-, second-, or third-order Markov chains (Carrera-Rivera et al., 2023). In the reported offline evaluation on a simulated industrial mixing machine app, the dataset contained over 10,000 interaction events and over 1,300 valid sequences, precision improved from approximately 32% to approximately 36% with higher-order models, recall remained above 88%, and mean reciprocal rank stabilized around 81% (Carrera-Rivera et al., 2023). Because recommendations are presented as minimally intrusive assistance that users can deactivate, the interface remains advisory rather than coercive.

Earlier adaptive HMI work in Product Data Management and Product Life Cycle Management framed similar goals through the I-SOAS conceptual architecture, especially its Intelligent User Interface component. Its design requirements are flexibility, agent-based design, context awareness, and natural communication mode. Concretely, this includes user-level detection, an automatic GUI trainer for beginners, drag-and-drop redesign, template saving and loading, multimodal input through mouse, keyboard, touch, and voice, and a Personal Assistant that remembers prior state and displays context-appropriate behavior (Ahmed, 2010). The prototype relies on Java and Rich Internet Application frameworks including Adobe Flex, AJAX, OpenLaszlo, and Silverlight (Ahmed, 2010). Although the paper does not formalize adaptation mathematically, it treats interface plasticity as an HMC requirement rather than a cosmetic feature.

In mixed traffic, the same design principles appear in vehicle-to-driver communication. A human-centered design process generated eight HMI concepts for autonomous-vehicle communication with conventional-vehicle drivers at intersections; the highest-rated concept was implemented both as an external HMI and as an internal HMI overlaid on the conventional vehicle’s windshield. In the reported VR study, both HMI conditions significantly increased situation awareness and trust relative to control, while the internal HMI yielded the highest situation awareness and the lowest increase in pupil diameter, and led to earlier and more pronounced braking when the autonomous vehicle insisted on right-of-way (Avetisyan et al., 2023). The broader point is that HMC often depends less on raw information quantity than on addressability, salience, and egocentric framing.

3. Embodied communication, haptics, and shared control

Embodied HMC treats force, timing, and motion as communicative media rather than mere side effects of control. Roche and Saint-Bauzel’s study of physical human-human interaction in co-manipulative tasks used a one-degree-of-freedom robotized setup with haptic-only communication and compared human-human dyads, human-robot dyads, and solo performance. Several physical predictors of shared choice were evaluated, but the “First Crossing” parameter—defined as the time at which either participant’s handle deviates by at least 30% of the target distance from center—achieved 94.2% prediction accuracy at $t=0.9$ s and an average prediction time of 0.899 s (Roche et al., 2016). The associated statistical state machine places the robot in a default follower mode, lets the human lead if initiative is detected, and otherwise switches the robot to a leader mode with a minimum-jerk trajectory; in 66% of human-robot trials the human took initiative, and in 34% the robot did (Roche et al., 2016). This work identifies initiative timing as a robust communication primitive for cooperative manipulation.

Automated driving extends the same principle to authority transfer. An intelligent two-phase haptic interface on the steering wheel switches between predictive guidance and haptic assistance according to the driver’s cognitive and physical state. When control ability is low or medium, the system provides guiding torque to help the driver recover state and control; when control ability is high and driver intervention exceeds 90% for 1.5 s, it shifts to minimal haptic assistance and then relinquishes control completely. The control-sharing model is summarized by $T_{ref} = T_H + T_A + T_{hpt}$, and the assistance phase by $T_{hpt} = C (T_{ref} - T_H)$ (Lv et al., 2020). In vehicle experiments with 26 participants, takeover time decreased from $8.0 \pm 0.6$ s to $4.4 \pm 0.2$ s in one task and from $7.9 \pm 0.8$ s to $4.4 \pm 0.3$ s in another, accompanied by lower variability in steering torque, steering angle, and yaw rate (Lv et al., 2020). Here communication is realized as adaptive torque, not symbolic dialogue.

Human-robot collaboration studies comparing modalities show how interface channel structure affects embodied work. A modular ROS-based architecture supports smartwatch-based gestural input and tablet-based touchscreen input using a finite-state machine that maps gesture “signals” to sequential transitions and touchscreen “events” to direct state jumps. Gesture recognition is implemented with the SLOTH online LSTM-based recognizer. The study reports high precision but low recall for gesture recognition, touchscreen superiority in perspicuity, efficiency, and dependability, and gesture superiority in stimulation and novelty; 23 of 25 participants completed all touchscreen tasks within the time limit, compared with 5 of 25 for gesture-based tasks (Bongiovanni et al., 2022). The comparison shows that communicative richness and hands-free interaction do not automatically translate into task efficiency when recognition recall is limited and navigation is sequential.

4. Neural, physiological, and affective channels

Recent HMC research increasingly treats internal human processes as communication signals in their own right. DS-MTNet addresses the observation that most current HMC systems rely on environment-facing sensors while the operator’s perceptual, intention-related, and state-related processes remain insufficiently integrated into machine perception. The framework jointly decodes three EEG-derived readouts—lane-departure-related epochs for environmental-event processing, steering-response stage for response preparation, and reaction-time-defined alertness state for internal state—using three streams: EEG waveforms, task-routed source embeddings, and temporal-spectral power features (Yu et al., 7 Jul 2026). Its routing is formalized at the source level by

$$\hat{S}_{m, r}(t) = \sum_{b=1}^{F} \sum_{k=1}^{K} a_{m,r,b,k} S_{b,k}(t),$$

and at the slot level by

$Z_q = \phi_q(\tau_q^T V).$

On a sustained-attention driving EEG dataset, the model achieved 81.70% ACC for lane departure, 85.12% ACC for steering response, and 77.60% ACC for alertness state, outperforming traditional, single-task deep, and multi-task baselines, with the most robust gains for steering-response stage decoding (Yu et al., 7 Jul 2026). Interpretability analyses further showed that each task predominantly used a unique information slot with approximately 70% preference (Yu et al., 7 Jul 2026).

Physiological HMC in heavy equipment operation grounds similar distinctions in EEG and EMG. The dual-loop task-allocation study on excavator operators measured EEG power spectral density in alpha and beta bands and corticomuscular coherence (CMC) during trenching and trench-finding. In the complex trench-finding task, experts showed higher alpha power in central and parietal regions, novices showed higher beta power in frontal and central areas, and experts showed significantly higher CMC in beta and gamma bands, especially in C3 and Cz, indicating stronger brain-muscle coupling (Xu et al., 2024). The coherence measure is

$t=0.9$0

No significant group differences were reported for the simpler trenching task (Xu et al., 2024). The evidence supports the claim that physiological markers can distinguish intuitive from intellectual operating modes in real-world HMC.

Affective state modeling addresses a different internal channel: communication failure itself. In two studies of voice-based HMI, tasks were designed to induce ASR failure and machine speech failure alongside non-communication puzzle and riddle tasks. The resulting annotations and self-reports identified “confused,” “frustrated,” “irritated,” “thoughtful,” “realization,” and related labels, and showed that confusion was the most frequent label during communication errors (Kim et al., 2022). In self-report data, 76% reported feeling confused, and 75% of those reports occurred during communication tasks; confusion scores were higher for failed tasks, with averages of 2.93 versus 1.67 for successful tasks (Kim et al., 2022). The study argues that confusion is not a single emotion with a unique facial signature, but a context-dependent affective complex that unfolds with frustration, irritation, uncertainty, or focused attention. This suggests that HMC systems require temporal, context-aware affect sensing rather than static emotion classification.

5. Semantic and wireless infrastructures

At the communication-theoretic level, HMC increasingly departs from Shannon-style bit transport toward meaning- and task-oriented exchange. Semantic communication distinguishes H2H, H2M, and M2M settings, with H2M defined as semantic techniques for conveying meanings understandable by both humans and machines so that they can interact, and M2M defined as effectiveness techniques for efficiently connecting machines such that they can execute a computation task in a wireless network (Lan et al., 2021). The proposed architecture adds a semantic layer inside the application layer and introduces cross-layer signals including Channel Rate Information, Data Importance Information, Partial Algorithm Information, and Data Type Information (Lan et al., 2021). Its task-oriented perspective is often formalized through the information bottleneck: $t=0.9$1 and the paper also emphasizes knowledge graphs as shared semantic background for encoding, decoding, and semantic error correction (Lan et al., 2021).

Wireless HMC extends this semantic focus into geographically distributed collaboration. A generic WHMC architecture comprises six tightly coupled components—human, interface, sensors, actuators, controller, and network—and can be organized by agent autonomy, agent multiplicity, and agent geography (Pang et al., 26 Feb 2026). To evaluate such systems, the literature introduces Quality of Collaboration,

$t=0.9$2

and identifies five communication requirements: accessibility, transparency, scalability, resilience, and sustainability (Pang et al., 26 Feb 2026). The same work highlights enabling technologies including Human-Bond Communications, semantic and goal-oriented communication, integrated sensing and communication, terahertz communication, reconfigurable intelligent surfaces, rate-splitting multiple access, soft sensors and actuators, and advanced HMI such as XR, tactile interfaces, and brain-computer interfaces (Pang et al., 26 Feb 2026). The emphasis on QoC indicates that latency, jitter, reliability, and bandwidth are relevant only insofar as they mediate collaborative task performance and human fluency.

A more formal WHMC model introduces dual wireless loops for machine and human control, short-packet transmission, fading channels, and advanced HARQ, while modeling human control lag as a finite-state Markov chain with state set $t=0.9$3, transition matrix $t=0.9$4, and stationary distribution $t=0.9$5 (Pang et al., 2024). The resulting stochastic cycle-cost analysis derives an explicit stability condition expressed in terms of wireless channel statistics, human dynamics, and control parameters, and validates it in simulations and a collaborative cart-pole experiment (Pang et al., 2024). In that experiment, collaborative control achieved better stabilization than human-only or machine-only control (Pang et al., 2024). The broader implication is that, in wireless HMC, communication delay and loss are not external nuisances but constitutive parts of the human-machine control loop.

6. Evaluation, misconceptions, and theoretical tensions

Evaluation remains a major unresolved issue in HMC because the field spans dialogue, control, physiology, and networked collaboration. In emergent communication research, four common properties recur across the literature: game environment, learning paradigm, interaction type, and theory of mind. The review distinguishes machine-centered emergent communication, where agents develop symbolic protocols from scratch, from human-centered emergent communication, where agents are grounded in human natural language and must coordinate with humans (Brandizzi, 2023). Across these traditions, evaluation uses metrics as different as topographic similarity, message entropy, mutual information, BLEU, negative log-likelihood, and human ratings, and the review states that there is no single universally accepted metric (Brandizzi, 2023). Open challenges include generalization and compositionality, language drift, scaling to complex and long-term contexts, ethics and human factors, underexplored non-verbal communication, and limited work on mixed human-AI teams (Brandizzi, 2023).

Several recurrent misconceptions are explicitly documented. In architecture selection for embedded AI, the belief that deeper models are necessarily more accurate was held by 68.8% of subjects without an explanatory UI, indicating that HMC design can correct model-selection errors rather than merely present outputs (Schöning et al., 2023). In mixed traffic, the absence of standardized communication protocols between autonomous and conventional vehicles is identified as a source of potential confusion, which is why the study calls for further research into standard HMC designs for AV/CV interactions (Avetisyan et al., 2023). These cases show that communicative failure often reflects representational mismatch or protocol ambiguity rather than lack of computational power.

At the theoretical level, pragmatics-oriented work argues that HMC with LLMs cannot be adequately described by traditional human-centered semiotic hierarchies. One paper proposes the HMC framework as a better fit for machine-involving communication and organizes the field along functional, relational, and metaphysical dimensions (Gvoždiak, 8 Aug 2025). The same work argues that Gricean, truth-evaluative pragmatics is poorly matched to predictive systems such as LLMs, whereas probabilistic pragmatics—especially the Rational Speech Act framework—better fits optimization-based behavior. It also identifies three forms of substitutionalism—generalizing, linguistic, and communicative—and introduces “context frustration” to describe the paradox of increased contextual input paired with collapse in contextual understanding (Gvoždiak, 8 Aug 2025). This suggests that future HMC theory may need to treat meaning as continually co-constructed under asymmetries between human pragmatic reasoning and machine prediction, rather than assuming that human communicative categories transfer unchanged to machines.