Intention Communication: Methods & Applications

• Intention communication is the explicit transmission of purpose, planned actions, and rationale to enable coordinated behavior across diverse agents.
• It spans low-level motion signals to high-level planning cues, employing modalities such as visual, auditory, and haptic channels for effective interaction.
• Frameworks in this domain merge communication with goals, prediction, and action selection, enhancing coordination in robotics, MARL, and other multi-agent systems.

Searching arXiv for papers on intention communication and closely related frameworks. Intention communication denotes the explicit or inferable transmission of purpose, anticipated action, rationale, or task-relevant semantic content so that another agent—human, robot, learning system, or network endpoint—can understand, predict, and coordinate with the sender. In industrial human–robot cooperation, it is defined as the structured communication by a robot of what it is doing, is about to do, and why; in goal-oriented communication theory, it means designing transmission around the end-goal rather than message reconstruction; in language-model pragmatics, it is the sender’s communicative intent, the Gricean “what was meant,” rather than surface form alone (Chiossi et al., 18 Jun 2025, Gutierrez-Estevez et al., 2022, Kwon, 3 Jul 2026).

1. Conceptual scope and operationalizations

Across the recent literature, intention communication is not a single mechanism but a family of formalisms that differ by domain. A recurring distinction is between low-level motion disclosure and richer communication of plans, goals, and reasons. In collaborative robotics, intention is explicitly not limited to the next motion; it spans immediate actions, short-term plans, and long-term goals and rationale, and is externalized through lights, sounds, speech, haptics, and related channels (Chiossi et al., 18 Jun 2025). In co-speech gesture generation, intention is modeled as high-level communicative function—such as emphasis, deixis, mental state, or process—rather than as rhythmic accompaniment to speech (Liu et al., 21 May 2025). In cooperative MARL, intention may be encoded as future state and reward, as intended current action, or as the objective function that defines an agent’s strategic preference (Ye et al., 2022, Fang et al., 2023, Chahine et al., 2022). In semantic communication, intention is a natural-language instruction that specifies which semantic region or task-relevant content should be transmitted (Ye et al., 7 Aug 2025, Jiang et al., 26 Apr 2026).

Domain	Operationalization of intention communication	Representative paper
Industrial HRI	What the robot is doing, is about to do, and why	(Chiossi et al., 18 Jun 2025)
Gesture generation	Latent communicative functions behind gesture	(Liu et al., 21 May 2025)
Cooperative MARL	Future outcome, intended action, or objective function	(Ye et al., 2022)
Semantic communication	User instruction defining task-relevant semantic content	(Ye et al., 7 Aug 2025)
LLM pragmatics	Sender’s communicative intent beyond literal surface form	(Kwon, 3 Jul 2026)

A foundational theme is that intention communication is inherently inferential. Human cooperative communication has been formalized as active inference over other minds, driven by an adaptive prior that mental states can be aligned; communicative action is then a policy for reducing uncertainty about hidden beliefs, goals, and intentions while producing shared understanding (Vasil et al., 2019). This perspective is echoed in contemporary LLM work, where communicative intent is treated as a first-class interpretability object and distinguished from surface content. The important empirical finding there is that models may represent the sender’s intent robustly while failing to act on it in default output behavior (Kwon, 3 Jul 2026).

2. Design spaces for human–agent and human–robot intention communication

A major line of work frames intention communication as a structured design problem. In industrial cobot collaboration, a multidimensional space is defined by three axes: transparency level, task abstraction, and modality. Transparency is grounded in Situation Awareness-based Agent Transparency, with SAT1 for current state or immediate intent, SAT2 for reasoning, and SAT3 for projection of future actions and rationale. Task abstraction is divided into operational, tactical, and strategic levels, and modalities include visual, auditory, haptic, and multimodal combinations (Chiossi et al., 18 Jun 2025). A closely related agent-human framework generalizes this structure as Transparency, Abstraction, and Modality, and applies it across bystander interaction, cooperative tasks, and shared control (Li et al., 23 Oct 2025).

This design-space literature rejects the assumption that intention communication is exhausted by low-level motion legibility. In both frameworks, higher-level rationale and projected future behavior are treated as essential when collaboration requires calibrated trust, anticipation, and coordination. Operational information is associated with immediate motions and safety states; tactical information with sequencing, synchronization, and turn-taking; strategic information with long-term goals, priorities, and trade-offs (Chiossi et al., 18 Jun 2025). This suggests that a flashing light, a short tone, and a spoken explanation are not redundant variants of the same signal but different coordinates in a structured space.

The same work also emphasizes that modality choice is not secondary. Visual channels provide spatial precision; auditory channels offer salience without requiring gaze; haptic channels are private and robust in noisy or visually cluttered environments. For safety-critical events, multimodal redundancy is preferred; for frequent non-urgent updates, subtler channels may be better; and strategic “why” information is better served by richer media such as speech or augmented-reality visualizations (Chiossi et al., 18 Jun 2025, Li et al., 23 Oct 2025). An important counterpoint is the persistent trade-off between transparency and overload: both design frameworks explicitly note that more transparency is not always better, particularly during high-workload phases.

3. Embodied signals: gesture, movement, gaze, and interface behavior

Several works study intention communication through embodied cues rather than abstract message passing. In gesture generation, the problem is reformulated as $p(G \mid S, I)$ rather than $p(G \mid S)$, with gesture motion conditioned on communicative intention $I$ in addition to speech $S$. The "Intentional-Gesture" framework builds the InG dataset by augmenting BEAT-2 with intention annotations, introduces H-AuMoCLIP for intention-aware audio–motion alignment, and constrains a multi-codebook motion tokenizer with a temporal cosine semantic loss. On the BEAT-2 benchmark, it reports an FGD of 0.256 for all speakers, compared with 0.446 for GestureLSM, and a user study reports Mean Opinion Scores of 3.76 for realness, 4.11 for synchrony, and 3.92 for smoothness (Liu et al., 21 May 2025). The underlying claim is not merely better realism, but that gestures become semantically meaningful realizations of communicative function.

Industrial robotics work arrives at a similar conclusion from a different direction. A small anthropomorphic proxy robot, ARMoD, is used as a communication layer for non-humanoid industrial machines such as forklifts, combining speech, gaze, and referential gestures. Lab studies with gaze tracking and motion capture report faster reactions and quicker localization of goal points and objects under multimodal interaction, while LLM-enhanced dialogue did not produce higher task efficiency than fully scripted interaction (Schreiter et al., 25 Feb 2025). The result is significant because it locates intention communication in synchronized multimodal behavior rather than in verbal content alone.

A service-robot study in hospitality examines intention communication to a group rather than an individual. Two factors were manipulated: visualization of the order on the robot screen and personalized movement trajectory/stopping location. The combination of personalized stopping and visualization yielded the highest group-wise order-delivery accuracy, 87%, compared with 37% for general stop without visualization, and it was also the most preferred condition, chosen first by 60% of participants (Hong et al., 2024). The same study found significant effects of customer role, showing that group members do not experience the same communicative behavior identically. A plausible implication is that intention communication in multi-user settings must address asymmetry of perspective, not only content clarity.

4. Multi-agent coordination, planning, and emergent or engineered protocols

In multi-agent systems, intention communication is commonly operationalized as communication about future behavior under partial observability. IEC, or Intention Embedded Communication, defines another agent’s intention as its next observation and reward, $x_{t+1}^j = (o_{t+1}^j, r_{t+1}^j)$, and uses a VAE-based Inferring Belief Module to decode messages into beliefs over those future outcomes. Communication and belief inference co-evolve, and the method is reported to learn about 50% quicker than MADDPG, while ablations show performance reductions of about 38% without the inferring belief module, 60% without communication, and 30% without hidden states (Ye et al., 2022). Here intention communication is explicitly future-oriented and predictive.

M2I2 pushes this further by coupling communication to masked state reconstruction and joint-action prediction. Messages are top-$k$ masked observations selected by a Dimensional Rational Network trained through meta-learning, and the integrated latent representation is optimized both to reconstruct the global state and to predict joint actions. The paper interprets this as improving the ability to anticipate teammates’ intentions and reports the highest communication efficiency among the compared baselines across Hallway, Predator-Prey, SMAC, and SMAC-Communication (Sun et al., 2024). This suggests that intention communication is not only about what is sent, but also about how received information is structured so that future behavior becomes inferable.

A direct comparison between learned and engineered communication in cooperative task allocation reaches a different conclusion. Learned Direct Communication generates messages and actions concurrently, whereas the engineered Intention Communication method uses an Imagined Trajectory Generation Module and a Message Generation Network to summarize a short-horizon latent trajectory. Under partial observability, the engineered approach reaches 99.9% success on a $10 \times 10$ grid and 96.5% on $15 \times 15$, compared with 30.8% and 12.2% for Learned Direct Communication and 0% for the no-communication baseline (Hill et al., 4 Aug 2025). The contrast is not between communication and no communication, but between emergent unstructured signaling and explicitly future-oriented intention sharing.

Two additional formulations show how domain constraints shape what counts as intention. In game-theoretic motion planning for autonomous driving, an agent’s intention is its objective function $J^i$, and communicated intention is treated as one hypothesis among alternatives in a discrete Bayesian filter that compares predicted and observed trajectories to update intention likelihoods (Chahine et al., 2022). In multi-order execution for finance, intention-aware communication is modeled as multi-round exchange of intended current actions $a_{t,k}^i$, where agents communicate and iteratively refine within-timestep trade intentions before execution (Fang et al., 2023). In both cases, intention communication is inseparable from coordination under coupled constraints.

5. Intention-oriented transmission, semantic communication, and networked systems

A distinct but related literature treats intention communication as a communication-theoretic problem: transmit what matters for the downstream goal, not what best reconstructs the source. "Learning to Communicate with Intent" formalizes this as end-to-end optimization of the transmitter, channel interface, and task decoder around task loss rather than reconstruction loss. In image classification, this improves performance relative to JSCC at low SNR, and in Atari BreakOut the paper reports that a JSCC strategy is not better than random action selection even at high SNRs, whereas the goal-oriented method approaches the upper bound even at low SNRs (Gutierrez-Estevez et al., 2022). The conceptual shift is from fidelity to effectiveness.

User-intention-driven semantic communication makes this shift explicit. In UIDSC, user intention is a natural-language instruction interpreted by a multimodal large model into a mask $p(G \mid S)$0 over the image, with the ROI $p(G \mid S)$1 becoming the object of reconstruction. A mask-guided attention module and channel-state embedding then preserve intent-relevant content under AWGN or Rayleigh fading. On a Rayleigh channel at 5 dB, the paper reports improvements over DeepJSCC of 8% in PSNR, 6% in SSIM, and 19% in LPIPS (Ye et al., 7 Aug 2025). Intention communication here is not between two deliberating agents but between a user, a semantic encoder, and a task-oriented receiver.

An edge–cloud architecture for AI glasses extends the same idea to wearable systems. The glasses act as an edge semantic agent; a server-side VLM infers user intention from low-resolution snapshots and historical prompts, then selects lightweight preprocessing such as OCR, Canny edge detection, or YOLO cropping before semantic transmission. Simulation results report more than 50% bandwidth reduction depending on the task while maintaining task performance, and the system exhibits graceful degradation at low SNRs (Jiang et al., 26 Apr 2026). This suggests that intention communication can function as an online control signal for adaptive resource allocation.

Transportation and human–machine interfaces provide further variants. In ETSI VAMs, intention sharing for cyclists is encoded as uncertainty ellipses derived from EKF-CTRV predictions, reducing computational complexity by an order of magnitude relative to trajectory vectors while maintaining constant message size and reliable multisecond prediction horizons on real GNSS trajectories (Valle et al., 29 Jun 2026). In brain-to-vehicle communication, EEG is treated as a direct intention channel for steering intent; a CNN classifies left, right, and straight intentions from minimally preprocessed EEG with 83.7% accuracy (Alavi et al., 8 Jan 2026). These works differ in mechanism, but both encode intention as future-oriented action information transmitted under channel and latency constraints.

6. Pragmatics, interpretability, and the problem of acting on represented intent

The most explicit recent account of communicative intent in AI treats it as a latent pragmatic variable inside LLMs. For recognize versus evaluate, a linear probe cleanly decodes the sender’s intent from default-pass hidden states across six models and four families, with probe accuracy reaching 1.00 on Qwen2.5-3B at deeper layers while a leave-phrasing-out bag-of-words baseline stays at 0.48 (Kwon, 3 Jul 2026). The same work finds strong transfer to pragmatically inferred intent and to the support-versus-help contrast. The notable claim is that representation is not the main failure point.

Behavior tells a different story. Some models still respond to recognize-intent messages with unsolicited critique or feedback. In the discard models, default honoring rates are 0.65 for Qwen2.5-3B, 0.60 for Qwen2.5-7B, and 0.57 for Llama-3.1-8B, even though probe accuracy is near ceiling; steering along a discriminative intent direction raises honoring to 0.98, 0.95, and 0.85, respectively (Kwon, 3 Jul 2026). The paper therefore distinguishes representation from readout: the readout lags the representation in depth, and model behavior depends on whether the internal intent feature is routed into output generation.

This finding links contemporary interpretability back to older theories of human communication. Under active inference, cooperative communication is intentional communication geared toward aligning mental states, and communicative policies are selected for their epistemic and pragmatic value in reducing uncertainty about others’ hidden states and creating shared understanding (Vasil et al., 2019). A plausible implication is that current LLMs already instantiate part of this inferential structure—robust internal representation of what was meant—without yet reliably solving the downstream policy problem of responding in a way that honors that intent.

Across domains, intention communication is therefore best understood not as a single interface primitive but as a layered problem involving representation, transmission, modality, timing, inference, and action selection. The recent literature converges on a common conclusion: effective systems must communicate or infer not only what is happening, but what is meant, what will happen next, and why. Where the literature diverges is over how that should be achieved—through multimodal design spaces, engineered trajectory summaries, emergent or belief-based protocols, semantic masking, or activation-level routing—but the unifying target is the same: making autonomous behavior legible enough for coordination without collapsing under ambiguity, overload, or misread intent.