Skip to Content

Wiki source code of Human-Agent Teamwork

Last modified by Laura Ottevanger on 2022/04/05 13:51
Show last authors
1 In this section, we will introduce you to our implementation of the HAT (Human-Agent Teamwork) concepts. We will describe what these HAT concepts are followed by an example of how we envision the concepts to work with WAF (Walking Assisting Friend, i.e. the name of our prototype).
2
3 == Joint Activities ==
4
5 The first aspect deals with joint activities. There are three key factors that make up any joint activity:
6
7 * Interpredictability
8 * Common Ground
9 * Directability
10
11 Interpredictability is embedded in WAF through the [[Guiding Human Back IDP (Interaction Design Pattern)>>https://xwiki.ewi.tudelft.nl/xwiki/wiki/sce2022group02/view/Main/Design%20Patterns/]]. In this pattern, WAF should be perceptive to a PwD straying away from their path in order to effectively guide the human back. The perception should therefore include knowledge about the user's position and how likely they are to walk away to appropriately activate this IDP. Conversely, the user should be aware of WAF's affordances such as that it can produce sound (e.g. act as an alarm) or that it can only walk on flat surfaces to keep expectations real.
12
13 Common Ground is established in WAF by means of an [[ontology>>https://xwiki.ewi.tudelft.nl/xwiki/wiki/sce2022group02/view/Main/Ontology/]] that displays WAF's knowledge of the PwD and other relevant people, locations and functions. As PwD may not recall information accurately, it is difficult to say whether they are aware of the (implicit) knowledge they share with WAF at all times. Intuitively, WAF's behaviour, expressed through its functionality, should be reliable and predictable to avoid throwing off any PwDs.
14
15 Directability is also accounted for in our [[IDPs>>https://xwiki.ewi.tudelft.nl/xwiki/wiki/sce2022group02/view/Main/Design%20Patterns/]] by ensuring that WAF is aware of the volatility of the PwD. WAF should not only be able to perceive a sudden change in wanting to walk from the user but also come across as determined and capable of making a decision swiftly if needed. Unfortunately, it was not possible to account for it in the prototype of WAF due to technical limitations of the MiRo to perceive the user at all. For instance, by means of a camera contained in the MiRo or sound of a microphone.
16
17 == Mental Models ==
18
19 The second aspect is modelling the inner beliefs and knowledge of both the WAF and the PwD about the world around them into their respective mental models. Our [[ontology>>https://xwiki.ewi.tudelft.nl/xwiki/wiki/sce2022group02/view/Main/Ontology/]] roughly captures the intersection of WAF's and PwD's models as an SMM (Shared Mental Model). A change of knowledge in the SMM should update the respective beliefs of each actor's mental model.
20
21 === Shared Knowledge Types ===
22
23 Most of the task-related knowledge is contained within the role of the caretaker with some parts attributed to WAF. Once WAF is set up by the caretaker for one specific person with dementia and their unique needs, WAF should be fully autonomous in guiding the PwD and entertaining them on their walk. Team-related knowledge is shared by the PwD and WAF to ensure the interpredictability of one another. Factual knowledge is important for the WAF to be able to carry out its functions and to proactively account for any change that may impede its functionality. The intentions are subsequently communicated to either the PwD or the caretaker by means of a change in movement or a message about the state of WAF to the caretaker.
24
25 == Situation Awareness ==
26
27 The third aspect is Situation Awareness (SA) and comprises of three stages:
28
29 * Perception
30 * Comprehension
31 * Projection
32
33 The SA of PwD is usually limited to perception and a bit of comprehension, but the latter is unreliable. This is due to the PwD's reduced cognitive ability and their tendency to become quickly disoriented when presented with sudden changes. As such, it is up to the SA of WAF to pick up on changes in the environment, understand their meaning and foresee how these changes develop in the near future. For example, a PwD is in a nauseous state where they are about to pass out and collapse on the floor, WAF should pick up the falling motion, by measuring a rapid increase in the accelerometer, understanding that could imply that the user is falling and will hit the ground in due time. This awareness can cause WAF to sound the alarm timely to inform caretakers of a medical emergency.
34
35 == Coordination ==
36
37 The fourth is about coordination between members of the same team.
38
39 Coordination can be viewed in two regards:
40
41 * Top-Down Coordination
42 * Bottom-Up Coordination
43
44 The former is most suited when there are executives who assign roles to each team member on how to act and what their respective responsibilities are. However, in our case, the latter may be more suited as both members (i.e. PwD and WAF) are on equal footing and neither strictly imposes a role for the other. In fact, they are acting synchronously. For instance, when the PwD stops walking, WAF should stop and when WAF guides, the PwD is to follow it. Therefore, bottom-up coordination would work better for our purposes.
45
46 == Task Management ==
47
48 The fifth and last aspect of Task Management is two-fold:
49
50 * Workload Harmonization
51 * Work Agreements
52
53 By means of the [[design patterns>>https://xwiki.ewi.tudelft.nl/xwiki/wiki/sce2022group02/view/Main/Design%20Patterns/]] and taking into account the other aspects above, we endeavour to find the optimal workload balance between WAF and PwD to harmoniously improve the overall collaborative performance. With that said, the value of privacy for the PwD must be taken into account when designing a system that is more perceptive of the cognitive and affective state of the PwD. Thus, work agreements need to be set where the PwD, primary caretakers and their loved ones can decide what information to share with third parties.
54
55
56
57