In this short paper, we consider scenarios in human-robot collaboration where the robot relies on deliberation to activate communicative actions. We claim that reasoning about the perception of these actions is a key, but often disregarded ingredient for successful communication, and propose a pre-theoretical model that accounts for this. We also discuss the problems that may arise if perception is neglected when reasoning about communication. ration.
In many situations, humans and artificial agents work together to perform joint tasks. Each of them may have their own strengths and weaknesses related to their capabilities. Communication can help the human and the agent in communicating their needs and help each other. For example, in a Human-Robot collaborative setting, the robots may have to act as communicators, and autonomously communicate with their human teammate when the need arises. Consider an assembly scenario where a robot and a human are both placing various pieces on a board. In some cases, pieces may be defective, which the robot can perceive. In this case, the robot has to inform the human about the defective pieces such that the human can replace/discard them. Due to the nature of the environment, which is potentially noisy, the robot may need to use other communication forms than speech signals. For instance, it may instead display a visual clue on a monitor, or move the defective object on a dedicated space as a way to inform the human. The robot therefore has to reason upon the communication actions to perform. In this paper, we discuss some perceptual aspects required for successful communication in such a collaborative setting, and claim that these aspects ought to be taken into account when reasoning about communication activities..
Deliberation is often considered a necessary ingredient for communication in purposeful human-agent
interaction [
namely, reasoning on the content of the message based on the communicator’s knowledge, interlocutor, environment, and context so that the communicator can achieve the communicative goal; How to communicate, which includes two dependent levels: inter-how and intra-how, based on the modality used to represent the message (Inter-How) and the expressiveness of the message (Intra-How); and When to communicate, namely, identifying the need to communicate (When-Need), and reasoning behind acting (When-Act). As an example, a robot may use reasoning to decide to ask a human operator to carry a given box (What), and to ask it now (When) using its voice interface (How).
CEUR
ceur-ws.org
Just performing the action does not guarantee that communication will take place unless the action is perceived. In this paper, we consider communication as the action being performed by the communicator and being perceived by the interlocutor. In the context of deliberative communication considered here, this means that when deciding to perform communicative actions, the agent may have to reason about the perceiving capabilities of the human before and during communication to ensure that these actions are perceivable. Interestingly, these capabilities may change dynamically: for example, they may be afected by environmental factors, like noise, or by the cognitive and physical state of the human, like fatigue or impairment. Depending on the capabilities, the human may best perceive actions done via touch, auditory or visual means, and the agent can choose what communicative action to perform accordingly. In our assembly scenario, the robot may decide that the environment is too noisy for a voice message to be perceived, and decide instead to post a visual message on a monitor (How) as soon as the human will turn toward it (When). After having performed the action, the human may or may not have perceived the action. If the action is not perceived, the message will not be conveyed to the interlocutor. Therefore, the agent also needs to monitor and reason about the uncertainty in the action being perceived by the human after the action is performed. All these reasoning activities require a model of communicative actions that include perception explicitly: this paper reports some initial reflections toward the development of such a model, where, before an artificial agent decides to communicate with a human, the action being perceived must also be considered.
In the field of Human-Robot Interaction (HRI), much attention has been put to the issues of
communicating with the human [
In the rest of this paper, we discuss the ingredients needed to reason about the perception of communication actions in a human-robot collaborative context, and discuss some of the challenges that arise when there is a mismatch between performing a communicative action, perceiving it, and knowing that it has been perceived.
In Speech Acts theory [
– “actions performed by an agent, with the intention of increasing another agent’s knowledge of
the first agent’s state of mind.”
Existing works have defined communicative actions in various contexts as:
• Hellström et al. [11]
• Frijns et al. [12]
• Sabu et al. [
The definition by Frijns et al. [12] does not distinguish between actions and communicative actions. Since we consider the need for deliberation in communication and with agents deciding to communicate, such a distinction is necessary. The definitions by Hellström et al. [11] and Frijns et al [12] consider the actions being perceived only implicitly. Both definitions focus on the goal which the communication aims to achieve, whereas the definition by Sabu
et al. [
In this work, we adopt the definition of communicative actions by Sabu et al. [
We now present our preliminary model, that focuses on dyadic communication between a robot agent ( ) as the communicator, and a human agent ( ) as the interlocutor. We assume the agents are deliberating and acting in a partially observable environment where actions can have non-deterministic efects. The model below is pre-theoretical, in that we do not propose a full formalization but only present those elements that are the focus of this paper.
Consider a pair of agents A = { , }
, and finite set of states where ⊆ × × . are the ontic components of states, representing the physical state of the environment and of the agents; are the epistemic components of states, representing the mental state of the robot and of the human, and respectively. A state ∈ is a triple ⟨ , , ⟩. Each agent ∈ A can perform a finite set of actions = ∪
∪ { 0}. is the set of ontic actions that can perform, i.e., those that result in a change in the ontic state, like the action MOVE_defective-obj. is set of epistemic actions that can perform, i.e., those that result in a change in the epistemic state, like the action SAY_defective-obj-name. Note that in cognitive science, the term “epistemic action” is often used as a synonym of information gathering action, that is, an action that only changes the epistemic state of the actor [13, 14]. In our model, we use the terminology from Dynamic Epistemic Logic [15] and postulate that an epistemic action can change the epistemic state of the communicator ( ), of the interlocutor ( ), or both. Lastly, 0 is the null action that does not produce any change in state, introduced for convenience.
An important feature of actions in our context is that actions may have communicative content. This is obvious for epistemic actions that change the epistemic state of the other agent. However, ontic actions can also have a communicative content and therefore afect the epistemic state: e.g., the previous action MOVE_defective-obj may convey the message that the object needs to be replaced. We refer to any action with a communicative content as a communicative action. The crucial observation here is that an agreement has been established between the agents on a shared meaning of these actions, which we informally refer to as their message. In our model, we assume that such an agreement exists, and that communicative actions are interpreted accordingly by both and . In our example, the robot and the human agree that moving an object to a certain dedicated space means that the object is defective.
We now come to the main point of this paper, that is, the claim that communicative actions must be perceived in order for their efect on the epistemic state of the interlocutor to materialize, and that any deliberative communication should take this fact into account. To model this, we introduce, for any agent ∈ A, the following predicate:
• Perceive ( , ) if perceives ∈ in performed by in state , where ∈ , ∈ A and ≠ .
We can use this predicate to give a more precise definition of the robot communicative actions
introduced by Sabu et al. [
A communicative action is an action ∈ representing a message, performed in a state ∈ , produces a change in provided that Perceive ( , ) is true. state of the human, provided that the human has perceived the action.1 could result in ′ = ⟨ ′, ′ , ′ ⟩, where ′ is a new state of the environment and In the previous example, the robot’s executing the ontic action MOVE_defective-obj in = ⟨ , ′ is a new epistemic , ⟩ In the previous example, performing also changed the epistemic state ′ of the performing agent . One obvious reason for this change is that may know that it has performed the action. In this paper, we are more interested in another reason why ′ may change: after performing a communicative action, the communicator may have the belief that the action has been perceived, which in turns implies the belief that its epistemic efects have been produced. To model this, we introduce, for any pair of agents , ∈ A with ≠ , the following predicate:
• Believe Perceive ( , ) if believes that perceives ∈ in performed by in , where ∈ . In general, if a communicative action ∈ is applicable in a state = ⟨ , , ⟩, then performing in changes to a new state ′ = (, ), where is a transition function.
How the epistemic components of are afected by the action depends both on Believe Perceive ( , ) , as summarized in Table 1. Note that how the ontic component ′ changes only depends on , but is independent on the action’s being perceived or not. To illustrate, consider our assembly scenario where the robot has to decide to communicate in state by either performing the action MOVE_defective-obj or SAY_defective-obj-name. If the environment is noisy, and the Believe Perceive (SAY_defective-obj-name, ) is False, the robot can decide to perform the action MOVE_defective-obj, provided that Believe Perceive (MOVE_defective-obj, ) is True. Under these conditions, the action will be perceived, which results in successful communication. Perceive ( , ) and on may entail a request to the interlocutor to perform an action. For instance, the robot’s (epistemic) action SAY_defective-obj-name may entail the request to the human to perform the (ontic) action to 1A possible variation is one where the message is conveyed by the efect of the action rather than by its mere performance: in that case the Perceive predicate should use the resulting state as an argument, but the essence of our discussion would not change. dispose of the object; the (ontic) action MOVE_defective-obj may entail the same request (assuming the right agreement, as we discussed above); and the (epistemic) action SAY_ask-obj-name may entail the request to perform the (epistemic) action to tell the object name to the robot. In our initial model, we assume that requests to the human will always be complied to: under this assumption, a communicative action done by the robot may result in additional, mediated efects, 2 namely, the efects produced by the action ( ) requested to (and executed by) the human. We refer to these efects as after-efects . Depending on the request contained in , the after-efects can be ontic, epistemic, or null. Also, these efects can be produced in the state ″, resulting from the execution of in state ′, or in a later state ‴, if the human postpones the execution of : we refer to these as immediate and delayed after-efects, respectively.
An important observation regarding the after-efects of a communicative action is that these efects can only materialize if Perceive ( , ) is true. Modeling this can be important when reasoning about communication.
Consider again the diferent cases in Table 1. Case 1 is the nominal case, in which the action performed as communication by results in the action being perceived by , and correctly believes that the action has been perceived. Hence, will expect that the states resulting from the execution of will be afected by both the efects of , and by its after-efects, as we discussed above. In our example, the robot would expect that after SAY_defective-object-name is executed, the object will be disposed of.
In case 4, the action performed by is not perceived and, correctly, does not believe that the action has been perceived. Thus, the action performed by the communicator does not result in the change of the mental state of , nor in the related after-efects. In our example, the object will not be disposed of, and the robot knows that it will not be disposed of: this would enable the robot to, e.g., decide to do the communication again, possibly using a diferent action.
Cases 2 and 3 result in false beliefs and may lead to problematic situations. These false beliefs may arise due to the partial observability of the actions and epistemic states of other agents. In case 2, although the action performed by was perceived by , has a false belief that it was not perceived. As a result may decide to perform again, which could result in annoyance on . In case 3, the action performed by is not perceived by , but has a false belief that is was perceived. As a result, may move on with execution of other actions that presume that the communication has taken place, or it would wait to observe the efects expected to be produced by . In our scenario, the robot might have used a SAY_defective-obj-name to communicate a defective object, unaware that the human can not hear that because of the noise level, and then wait forever that the human replaces the object.
Note that the above four cases require diferent decisions by the robot after has been performed. They may also lead to diferent decision about which to perform in the first place, depending on the context. These cases of course could not be distinguished if we did not explicitly represent and reason about perception.
In this work, we discussed the importance of reasoning about communicative actions by a robot collaborating with a human, which — crucially — includes reasoning about whether the communicative action was perceived by the human. Two situations were highlighted in which the robot can have false beliefs where the action may or may not be perceived by the human. As a result, the robot may unnecessarily perform the same action again, or erroneously move on with the execution of other actions, or wait for efects that will not be produced. To address this, the robot has to reason about the uncertainty in the action being perceived by the human. For simplicity, we considered the robot 2This relates to the ramification problem , which has been extensively discussed in the literature [16]. agent (communicator) performing actions representing a message with the human agent (interlocutor) having to perceive the action. But communication can also be considered as having the message being perceived by the interlocutor instead of the action. Also, while we focused our discussion on human-robot communication, it is worth mentioning that similar considerations (and solutions) could be made for any agent-agent communication, whenever the reliability of communication can be afected by contextual conditions.
The pre-theoretical model presented in this work is not formal. Future work will focus on formalizing the states and actions within the scope of deliberative communication for human-robot collaboration, and use this formalization to address the challenges with respect to the false beliefs of the perceiving of actions. Future models will also address possible delays in perceiving the message. Another challenge that will be addressed is the uncertainty in the human producing the after-efects of communicative actions when the assumption of human complying to a request is relaxed, as in the real-world there is no guarantee that this assumption will always hold.
This work was funded by the Swedish Research Council, grants number 2022-04676 and number 202105542, and by the European Commission, under the Horizon AI4Europe project, grant agreement 101070000.
[11] T. Hellström, S. Bensch, Understandable robots - What, Why, and How, Paladyn, Journal of
Behavioral Robotics 9 (2018) 110–123. doi:10.1515/pjbr-2018-0009. [12] H. A. Frijns, O. Schürer, S. T. Koeszegi, Communication Models in Human–Robot Interaction: An Asymmetric MODel of ALterity in Human–Robot Interaction (AMODAL-HRI), International Journal of Social Robotics 15 (2023) 473–500. doi:10.1007/s12369-021-00785-7. [13] D. Kirsh, P. Maglio, On distinguishing epistemic from pragmatic action, Cognitive science 18 (1994) 513–549. doi:https://doi.org/10.1016/0364-0213(94)90007-8. [14] S. Croom, H. Zhou, C. Firestone, Seeing and understanding epistemic actions, Proceedings of the National Academy of Sciences 120 (2023) e2303162120. doi:https://doi.org/10.1073/pnas. 2303162120. [15] T. Bolander, M. B. Andersen, Epistemic planning for single-and multi-agent systems, Journal of
Applied Non-Classical Logics 21 (2011) 9–34. doi:https://doi.org/10.3166/jancl.21.9-34. [16] M. Shanahan, The ramification problem in the event calculus, in: Proceedings of the 16th International Joint Conference on Artifical Intelligence - Volume 1, IJCAI’99, Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 1999, p. 140–146. doi:10.5555/1624218.1624239.