Implicit communication plays a crucial role in human collaboration, where contextual cues, such as intentions and situational awareness, foster shared understanding and seamless coordination. However, translating these capabilities to robots in collaborative tasks remains challenging. This work investigates the role of implicit communication in human-robot teams through three integrated studies. First, we analyse how linguistic implicatures influence performance and user experience in human-robot physical collaboration. Second, we explore how multi-modal implicit cues enhance team efectiveness and alignment when used proactively and reactively by robots. Building on these findings, the final phase focuses on designing and evaluating a multi-LLM robotics system that learns and adapts human teammates' communication strategies. By advancing robots' capacity to interpret and generate contextually appropriate implicatures, this research bridges gaps in human-robot collaboration, enabling the natural establishment, maintenance, and dynamic adaptation of common grounds.
CEUR Workshop
ISSN1613-0073
Please also sort these cubes.
Can you move it to here please?
You could take the small yellow cylinder ...
Let's move on to the third column.
And the last one.
The blue block needs to .... Oh, actually. Wait.
instance, the statement “I’m cold” can imply a request to close the window, depending on the situation.
ISAs exemplify how humans eficiently establish common ground, as they rely on mutual knowledge
and situational context while enabling listeners to actively infer unstated goals. Studies suggest
humans tend to use implicit ways to convey intentions when interacting with robots at frequencies
similar to those used with other humans [
Prior work in HRI has focused on implicit human-to-robot communication [
Building on the understanding of implicit communication, we aim to enhance robotic systems by
integrating multimodal implicit cue interpretation and generation, enabling robots to dynamically
adapt to evolving common ground. By leveraging multimodal large language models (LLM) that infer
diferent modalities such as speech, gaze, and motion, robots can potentially improve their ability to
understand and respond to implicit intentions in a more nuanced manner. Additionally, by adapting to
human behaviours and learning to generate appropriate implicit cues themselves, robots may become
more adept at facilitating natural, bidirectional communication. However, relying on a single LLM
presents several challenges. Due to knowledge gaps or misinformation, LLM outputs may sufer from
hallucination or bias [
In summary, the contribution of this research lies in enhancing human-robot collaboration through implicit communication, grounded in Clark’s theory of common ground, by implementing a threephase approach. We present preliminary findings from the first phase, demonstrating how robots’ interpretation of ISAs improves task eficiency and user experience in collaborative manipulation tasks. A full analysis and discussion of these results are detailed in our full paper [28]. Future work will extend to bidirectional implicit communication (phase 2) and adaptive multi-LLM systems (phase 3), aiming to refine robots’ ability to sustain and evolve common ground in dynamic settings. By bridging the gap between rigid, explicit robot communication and the fluidity of human teamwork, this research paves the way for cobots that integrate seamlessly into real-world collaborative environments.
This research is based on lab empirical studies that involved a robot and a participant. We use TIAGo (Figure 2(a)), a mobile manipulator robot with anthropomorphic features, including a 2 Degrees of Freedom head, neck, torso, and 7 Degrees of Freedom arm. TIAGo is equipped with an RGB-D camera, stereo microphone, and speaker that is suitable for HRI research and tabletop manipulation tasks [29]. We built a motion teleoperation interface and speech control software using the Robotics Operating System (ROS) and the TIAGo API. The arm actions are implemented by inverse kinematics, which calculates the joint configuration based on the desired Cartesian coordinates of the end efector [ 30]. These customised tools can facilitate experiments to control the robot in pilots and user studies.
For the first phase, we conducted a lab study ( Figure 1), engaging a participant and a robot in three collaborative physical tasks, to investigate the impacts of linguistic implicatures on HRC. Specifically, we aimed to answer these research questions: (RQ1) How does a robot’s capability to understand indirect speech acts influence the perceived team’s performance? (RQ1.1) How does a robot’s capability to understand indirect speech acts influence the fluency of human-robot teamwork? (RQ1.2) How does a robot’s capability to understand indirect speech acts influence the establishment of goal alignment among the human-robot team? (RQ2) How does a robot’s capability to understand indirect speech acts influence a human teammate’s trust in the robot’s performance? (RQ3) How does a robot’s capability to understand indirect speech acts influence a human teammate’s perception of the robot’s anthropomorphism?
The experiment involved 36 participants, who were divided into two groups: one interacting with a robot capable of interpreting ISAs and another interacting with a robot that could only respond to direct commands. Each participant engaged in three tasks: assembly, sorting, and polishing. This was a Wizard-of-Oz (WoZ) study and employed a mixed-methods approach, using both quantitative data (team fluency [ 31], goal alignment [31], performance trust [32], and robot anthropomorphism [33] questionnaires) and qualitative data (semi-structured interviews). Cumulative Link Mixed Models (CLMMs) via the ”ordinal” package in R [34] have been used for quantitative data analysis. This analysis is appropriate given the ordinal nature of dependent variables, while also allowing us to account for efects for each participant, the type of tasks, and each sub-item of the used scales. The interview results were transcribed and analysed using thematic analysis [35] by two researchers. After individually coding the transcripts and identifying initial themes, they discussed any discrepancies and refined the themes through consensus. The experiment setup and the procedure are shown in Figure 2.
Our findings demonstrate that the robot’s capacity to interpret linguistic implicatures plays a pivotal role in establishing common ground during HRC, although its efectiveness can vary depending on the context. The quantitative results show the robot’s ability to comprehend implicatures significantly enhances participants’ perceived team fluency, goal alignment, trust, and anthropomorphism Table 1. Moreover, this alignment of understanding fosters a sense of joint agency, where humans perceive the robot not as a passive tool but as an active partner engaged in co-constructing shared context.
Other qualitative results show that participants felt more confident in having established common ground with the robot when they successfully used ISAs to communicate. These indirect requests often emerged subconsciously, reflecting a natural inclination to leverage shared context, such as prior commands, environmental afordance (e.g., object locations), and multimodal cues (e.g., gestures). However, the efectiveness of ISAs can be task- and context-dependent. While indirect requests enhanced trust in ideal scenarios, participants expressed concerns about error-prone real-world settings, where misinterpretations could lead to mistrust.
These insights highlight the inherent limitations of relying solely on explicit command-based interactions, which lack the subtlety required for establishing shared understanding and the sense of teaming. The results also emphasise the importance of using implicit communication in a contextually adaptive and appropriate manner. Therefore, the careful integration of explicit information and implicatures in verbal communication emerges as a critical factor in optimising the performance and overall experience of human-robot collaboration.
Based on the findings from phase one, where humans and robots established common ground by linguistic implicatures, phase two investigates how robots can maintain shared understanding through bidirectional implicit communication in physical collaborative tasks. We evaluate three interaction paradigms: 1) when the robot proactively uses implicatures to convey contextual information; 2) when the robot uses implicit cues as backchanneling to respond to human-implied intentions. Specifically, we compare four methods of communication: explicit speech, implicit speech, explicit motion, and implicit motion, across two scenarios: human-lead (robot backchannels) and robot-lead (robot initiates). Metrics include team fluency, eficiency, goal alignment, and trust. Overall, this phase aims to identify how implicit cues maintain common ground in physical collaborative tasks and how mismatched communication disrupt collaboration.
Phase three integrates insights from phases one and two into designing and evaluating a robotic system that employs cooperative multimodal multi-LLMs for human-robot collaborative manipulation tasks. In this system, each LLM will be initialised from a pre-trained base model and progressively ifne-tuned for distinct roles, such as interaction perceiver, interaction generator, environment perceiver, task planner, and executor. The LLMs will cooperate to generate responses for each role, mitigating issues such as hallucinations. A user experiment will compare the performance of the multi-LLM system with that of a single LLM system on a simple pick-and-place task (low ambiguity) and a complex manipulation task (dynamic context shifts). The system evaluation will include both quantitative metrics, such as response time and task success rate, and qualitative analysis of failure cases. Additionally, we will assess the interaction through quantitative measures, including team fluency, eficiency, goal alignment, and trust, alongside qualitative insights into human perception of how efectively the system adapts to common ground. This phase bridges the iterative establishment and maintenance of shared understanding with real-time adaptation, advancing robots from rigid collaborators to partners capable of fluid, human-like coordination.
This research enhances human-robot collaboration by integrating implicit communication through three phases, guided by Clark’s common ground theory. By advancing robots’ ability to establish, maintain, and dynamically adapt shared understanding through multimodal implicit cues, we aim to enable more natural and context-aware teamwork. Our current findings demonstrate that implicit communication fosters the establishment of common ground, improving both team performance and user experience. Future work will explore how robots can maintain alignment using multimodal cues and dynamically adapt their strategies through multi-LLM architectures, ensuring robust collaboration in complex realworld environments. This research aims to transition HRC from rigid, direct communication to natural, nuanced, yet efective communication, ensuring adaptability and inclusivity for diverse populations in dynamic real-world environments.
During the preparation of this work, the author(s) used ChatGPT and Grammarly to check grammar and spelling. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content. arXiv:2305.08283 (2023). [23] S. Feng, W. Shi, Y. Wang, W. Ding, V. Balachandran, Y. Tsvetkov, Don’t hallucinate, abstain: Identifying llm knowledge gaps via multi-llm collaboration, arXiv preprint arXiv:2402.00367 (2024). [24] Y. Qin, S. Liang, Y. Ye, K. Zhu, L. Yan, Y. Lu, Y. Lin, X. Cong, X. Tang, B. Qian, et al., Toolllm: Facilitating large language models to master 16000+ real-world apis, arXiv preprint arXiv:2307.16789 (2023). [25] A. Zeng, M. Liu, R. Lu, B. Wang, X. Liu, Y. Dong, J. Tang, Agenttuning: Enabling generalized agent abilities for llms, arXiv preprint arXiv:2310.12823 (2023). [26] G.-C. Kang, J. Kim, J. Kim, B.-T. Zhang, Prograsp: Pragmatic human-robot communication for object grasping, in: 2024 IEEE International Conference on Robotics and Automation (ICRA), IEEE, 2024, pp. 3304–3310. [27] W. Shen, C. Li, H. Chen, M. Yan, X. Quan, H. Chen, J. Zhang, F. Huang, Small llms are weak tool learners: A multi-llm agent, arXiv preprint arXiv:2401.07324 (2024). [28] Y. Zhang, T. S. Ratnayake, C. Sew, J. Knibbe, J. Goncalves, W. Johal, Can you pass that tool?:
Implications of indirect speech in physical, arXiv preprint arXiv:2502.11720 (2025). [29] J. Pages, L. Marchionni, F. Ferro, Tiago: the modular robot that adapts to diferent research needs, in: International workshop on robot modularity, IROS, volume 290, 2016. [30] S. Chitta, E. Marder-Eppstein, W. Meeussen, V. Pradeep, A. R. Tsouroukdissian, J. Bohren, D. Coleman, B. Magyar, G. Raiola, M. Lüdtke, et al., ros_control: A generic and simple control framework for ros, Journal of Open Source Software 2 (2017) 456–456. [31] G. Hofman, C. Breazeal, Efects of anticipatory perceptual simulation on practiced human-robot tasks, Autonomous Robots 28 (2010) 403–423. [32] B. F. Malle, D. Ullman, A multidimensional conception and measure of human-robot trust, in:
Trust in human-robot interaction, Elsevier, 2021, pp. 3–25. [33] C. Bartneck, Godspeed questionnaire series: Translations and usage, in: International Handbook of Behavioral Health Assessment, Springer, 2023, pp. 1–35. [34] R. H. B. Christensen, ordinal—regression models for ordinal data, R package version 10 (2019) 54. [35] V. Braun, V. Clarke, Thematic analysis., American Psychological Association, 2012.