• No results found

Chapter 7 General discussion

7. Summary, Contributions & Limitations

7.5. Future directions

To summarise, lessons learned through the replication crisis in psychology and neuroscience, and the subsequent increased implementation of open science practices, may be of great value to the interdisciplinary HRI research community at large (Belpaeme, 2020). As various researchers have noted, conducting

embodied HRI experiments in conjunction with using brain imaging tools is not a trivial challenge – various pieces of hardware and software have to be

synchronised to ensure a smooth experimental procedure (Belpaeme, 2020; Perez-Osorio et al., 2018; Strait et al., 2020). Belpaeme (2020) foresees a new future in which experiments will benefit from increased rigor and the field might move towards more transparency in reporting findings – including ‘failed’

replications and null results.

7.5.1. Valuable tools (or: Where to go from here?)

Going forward, researchers working at the intersection of social robotics,

experimental psychology and cognitive neuroscience should strive to pre-register their hypotheses, study designs and analysis plans. Several platforms are

available for this purpose. Compared to the AsPredicted format, which is designed to cover essential aspects of the experimental design and analysis in nine short questions (e.g.: https://osf.io/ky4b7/), OSF preregistrations allow for more detail and nuance, especially when specifying mixed effects models (e.g:

uptake of Bayesian analysis methods across all fields (Baxter et al., 2016; Belpaeme, 2020; Shrout & Rodgers, 2018). Although more diverse analytical approaches are commended, researchers also point out that analysts should follow a “principled Bayesian workflow” (Schad et al., 2020, p. 1). For example, Bayesian modelling, which was described in Chapter 4, offers the researcher great flexibility regarding the types of models that can be specified and the kind of data that can be modelled. Despite accessible R packages, like ‘brms’,

beginners may struggle to navigate across the often-confusing labyrinth of

decisions that need to be taken in modelling (Bürkner, 2016; Schad et al., 2020). Here the authors again emphasize the crucial need for specifying models that describe a maximal effects structure already in the pre-registration, to limit later researcher degrees of freedom in the analysis. With greater acceptability of the publishing of null results will come a greater need for reporting evidence for the null. Although the modelling approach used here offers great flexibility, in many cases free programmes such as JASP might be a more beginner-friendly first step in Bayesian analysis (Belpaeme, 2020).

One promising opportunity for researchers at the intersection of these disciplines lies in new article formats (and journals), that place a focus on methodological rigor and meta-scientific perspectives. The vicious cycle of a decreased reward structure to pursue replication projects may be broken by Registered Reports (Chambers et al., 2015) and Registered Replication Reports (RRR) formats (Simons et al., 2014), which encourage large scale collaborative efforts and ensure acceptance at the journal before the results are known (Schimmack, 2020; Shrout & Rodgers, 2018). Adopting these new article formats may contribute to an overall more accurate picture of the scientific evidence base, as it has been recently revealed that the reporting rate of significant results is reduced from about 90 to a mere 50% when the Registered Report format is used (Scheel et al., 2020).

Importantly, these strides to encompass greater transparency in research

methods should ideally be accompanied by an acknowledgement and awareness of systematic disadvantages underlying large parts of the existing literature (Dworkin et al., 2020; Pownall et al., 2020; Zurn et al., 2020). In a recent preprint, we argue that just as the open science movement has prompted researchers to adopt more transparent approaches to research, feminist

psychology has much to contribute in constructing an equitable movement towards open science (Pownall et al., 2020). One small step could be to include citation diversity statements (Zurn et al., 2020), as implemented in Chapter 6 of this thesis. Whilst not completely without problems, this new convention may contribute to more transparency about diversity issues evident across scientific disciplines.

7.5.2. Transparent data visualisation

Another important tool for future HRI studies will be to use transparent data visualisation to communicate research findings (Allen et al., 2019). Repeatedly, HRI researchers have lamented issues with the interpretability of psychology and neuroscience findings, noting a missing shared language between disciplines (Baxter et al., 2016; Belpaeme, 2020; Irfan et al., 2018). One crucial factor in efficient communication of experimental findings is the use of clear data visualisations. Currently the most commonly used visualisations across research and the news media remain bar charts, although it has been shown that these visualisations lead to poor decision making when interpreting experimental findings (Newman & Scholl, 2012). For example, Newman and Scholl (2012) found that when they presented bar graphs, participants were more prone to believe that the data were contained within the bars. This visualisation method thus does not offer a good impression of the often-chosen measures of central tendencies. A popular new visualisation method - the raincloud plot – may be a better approach. These graphs depict raw data, distributional information and a boxplot with the median and interquartile range (Allen et al., 2019).

Throughout this thesis, I have used various types of visualisation methods that offer the advantage of “inference at a glance” (Allen et al., 2019, p. 33), including pirate plots (Chapter 3), box plots, density graphs of distributions (Chapter 4) and raincloud plots (Chapter 6). In his chapter on ‘fair statistical communication’, Dragicevic (2016, p. 291) applies “End User Dissatisfaction” (p. 311) as a metaphor. He critically reflects that the field of human-computer interaction (adjacent and interrelated with human-robot interaction, see Figure 1), despite its strong tradition of user experience studies, has adopted

visualisations that lead to suboptimal communication of empirical findings. New approaches to graphically representing data may serve as better “user interfaces

meant to help researchers in their task of producing and disseminating

knowledge, [and] the fields of HCI and infovis can take a head start and show the way to other disciplines.” (Dragicevic, 2016, p. 326)

7.5.3. Our future with social robots: beyond ‘the social brain’?

Recent years have not only seen a shift in methodological and analytical

approaches, but also a move towards incorporating new theoretical perspectives on social cognition (Cross & Ramsey, under review). As these authors argue, the current perspective of using a predominantly anthropocentric approach to investigating interactions with robots may limit the scope of potential questions and might overall stifle progress in this research area. Cross and Ramsey (under review) urge researchers at the intersection of psychology, neuroscience and social robotics to consider a wider, shared feature space between social agents (like humanoid robots) and objects, rather than focusing solely on the

commonalities and differences of processing human social interactions compared to interactions with machines. Overall, the authors argue that a more domain- general understanding of human cognition should be adopted, which echoes recent critical reflections by researchers requesting more nuance when parsing ‘the social brain’ – as it remains to be investigated how the purported specificity for social perception may be represented in the brain (Lockwood, 2020). These recent criticisms can also be seen as a challenge of Chevallier’s Social Motivation Theory (2012).

Figure 27 - Synthesis of the proposed stepwise process.

To summarise, the future of adapting paradigms for HRI may include new methodological, analytical and theoretical approaches, that will contribute to our understanding of sharing a social sphere with artificial agents. Through Figure 27, I have attempted to integrate the messages of several researchers working at the intersection of psychology, neuroscience and HRI, as well as one of the main conclusions from this thesis: before conducting studies with

humanoid robots, researchers should aim to conduct a (pre-registered) direct replication of the effect first (Irfan et al., 2018), then implement and adapt the paradigm appropriately for (embodied) interactions robots (Perez-Osorio et al., 2018), take these paradigms outside of controlled lab environments into the real world (Henschel, Hortensius, et al., 2020b; Pinti et al., 2018), and finally utilise the knowledge gained to inform the design of social robots (Wiese et al., 2017). These longitudinal and interdisciplinary efforts, as I have argued, are a

fundamentally challenging, yet ultimately rewarding undertaking, which may herald a sustainable future for social robots as the helpful companions we envision them to be.

Related documents