Autonomous robots can be dificult to understand by their developers, let alone by end users. Yet, as they become increasingly integral parts of our societies, the need for afordable easy to use tools to provide transparency grows. The rise of the smartphone and the improvements in mobile computing performance have gradually allowed Augmented Reality (AR) to become more mobile and afordable. In this paper we review relevant robot systems architecture and propose a new software tool to provide robot transparency through the use of AR technology. Our new tool, ABOD3-AR provides real-time graphical visualisation and debugging of a robot's goals and priorities as a means for both designers and end users to gain a better mental model of the internal state and decision making processes taking place within a robot. We also report on our on-going research programme and planned studies to further understand the efects of transparency to naive users and experts.
• Human-centered computing → Mixed / augmented reality; Systems and tools for interaction design; • Computing methodologies → Artificial intelligence ; • Software and its engineering → Software creation and management; Software design engineering; • Social and professional topics → Computing / technology policy. robots, mobile augmented reality, transparency, artificial intelligence
The relationship between transparency, trust, and utility is a complex one. By exposing the inner ‘smoke and mirrors’ of our agents, we risk of making them look less interesting. Moreover, the wide range of application domains for AI and of the diferent stakeholders interacting with intelligent systems should not be underestimated. Therefore, What is efectively transparent varies by who the observer is, and what their goals and obligations are. There is however a need for design guidelines on how to implement transparent ∗We thank the EPSRC grant [EP/L016540/1] for funding Rotsidis and Theodorou. Both authors contributed equally to the paper.
IUI Workshops’19, March 20, 2019, Los Angeles, USA
© 2019 Copyright for the individual papers by the papers’ authors. Copying permitted
for private and academic purposes. This volume is published and copyrighted by its
editors.
systems, alongside with a ‘bare minimum’ standardised
implementation [
Still, the use real-time implementation can help users to calibrate
their trust in the machine [13, and references therein]. Calibration
refers to the correspondence between a person’s trust in the
system and the system’s capabilities [
Yet, despite its efectiveness, there is a major disadvantage with the ABOD3 solution: a computer and display is required to run the software. One solution might be to port ABOD3 to run directly on robots with built-in screens, such as SoftBank Robotic’s Pepper. Albeit that this is a technologically feasible and potentially interesting approach, it also requires that custom-made versions of ABOD3 will need to be made for each robotics system. Moreover, this is not a compatible solution for robots without a display.
Nowadays, most people carry a smartphone. Such mobile phones
are equipped with powerful multi-core processors, capable of
running complex computational-intensive applications, in a compact
package. Modern phones also integrate high-resolution cameras,
allowing them to capture and display a feed of the real world.
That feed can be enhanced with the real-time superimposition of
computer-generated graphics to provide Augmented Reality (AR)
[
In this paper we demonstrate new software, ABOD3-AR, which can run on mobile phones. ABOD3-AR, as its name suggests, uses a phone’s camera to provide AR experience by superimposing the ABOD3’s tree-like display of Instinct plans over a tracked robot. 2
In this section we describe in some detail the tools and technologies used in our transparency experiments. 2.1
Behaviour Oriented Design is a cognitive architecture that provides
an ontology of required knowledge and a convenient representation
for expressing timely actions as the basis for modular
decomposition for intelligent systems [
BOD helps AI developers as it provides not only an ontology,
addressing the challenge of ‘how to link the diferent parts together’,
but also a development methodology; a solution to ‘how do I start
building this system’. It includes guidelines for modular
decomposition, documentation, refactoring, and code reuse. BOD aims to
enforce the good-coding practice ‘Don’t Repeat Yourself’, by
splitting the behaviour into multiple modules. Modularisation makes
the development of intelligent agents easier, faster, reusable and
cost eficient. Behaviour modules also store their own memories,
e.g. sensory experiences. Multiple modules grouped together form
a behaviour library. This ‘library’ can be hosted on a separate
machine, for example in the cloud.The planner executing within the
agent is responsible for exploiting a plan file; stored structures
describing the agent’s priorities and behaviour. This separation of
responsibilities into two major components enforces further code
reusability. The same planner, if coded with a generic API to
connect to a behaviour library, can be deployed in multiple agents,
regardless of their goals or embodiment. For example, the Instinct
planner has been successfully used in both robots and agent-based
modelling, while POSH-Sharp has been deployed in a variety of
computer games [
POSH planning is an action-selection system introduced by Bryson
[
Instinct is a lightweight alternative to POSH, incorporating
elements from the various variations and modifications of POSH
released over the years [
ABOD3 is a substantial revision and extension of ABODE (A BOD Environment), originally built by Steve Gray and Simon Jones. ABOD3 directly reads and visualises POSH, Instinct, and UN-POSH plans. Moreover, it reads log files containing the real-time transparency data emanating from the Instinct Planner, in order to provide a real-time graphical display of plan execution. Plan elements are highlighted as they are called by the planner and glow based on the number of recent invocations of that element. Plan elements without recent invocations dim down over a user-defined interval, until they return to their initial state. This ofers abstracted backtracking of the calls, and the debugging of a common problem in distributed systems: race conditions where two or more subcomponents constantly trigger and interfere with or even cancel each other. ABOD3 is also able to display a video and synchronise it with the debug display. In this way it is possible to explore both runtime debugging and wider issues of AI Transparency.
The editor provides a user-customisable user interface (UI) in line
with the good practices for transparency introduced by Theodorou
et al. [
ABOD3-AR builds on the good practice and lessons learned through the extended use of ABOD3. It provides a mobile-friendly interface, facilitating transparency for both end users and experts. In this section, we not only present the final system, but also look at the technical challenges and design decisions faced during development. 2.4.1 Deployment Platform and Architecture. The Android Operating System (OS) 1 is our chosen development platform. Due to the open-source nature of the Android operating system, a number of computer vision and augmented reality (AR) libraries exist. Moreover, no developer’s license is required to prototype or release the final deliverable. Android applications are written in Java, like ABOD3, making it possible to reuse its back-end code. Unlike the original ABOD3, ABDO3-AR is aimed exclusively for embodiedagents transparency. At the time of writing, Instinct (see Section 2.2) is the only supported action-selection system.
Our test configuration, as seen in Figure 1, includes the triedand-tested R5 robot. In the R5 robot,the callbacks write textual data to a TCP/IP stream over a wireless (WiFi) link. A JAVA based Instinct Server receives this information, enriches it by replacing element IDs with element names and filters out low-level information, 1https://www.android.com/ sending this information any mobile phones running ABOD3-AR. Clients do not necessarily need to be on the same network, but it is recommended to reduce latency. We decided to use this ‘middleman server’ approach to allow multiple phones to be connected at the same time. 2.4.2 Robot tracking. Developing an AR application for a mobile phone presents two major technical challenges: (1) managing the limited computational resources available to achieve suficient tracking and rendering of the superimposed graphics, and (2) to successfully identify and continuously track the object(s) of interest. 2.4.3 Region of Interest. A simple common solution to both challenges is to focus object tracking only within a region of the video feed, referred to as the Region of Interest (ROI), captured by the phone’s camera. It is faster and easier to extract features for classification and sequentially track within a limited area rather than over the full frame. The user registers an area as the ROI, by expanding a yellow rectangle over the robot. Once selected, the yellow rectangle is replaced by a single pivot located at the centre of the ROI. 2.4.4 Tracker. Various solutions were considered; from the builtin black-box tracking of ARCore 2 to building and using our own tracker. To speed-up development, we decided to use an existing library BoofCV 3, a widely-spread Java library for image processing and object tracking. BoofCV was selected due to its compatibility with Android and the range of trackers available for prototyping.
BoofCV receives a real-time feed of camera frames, processes
them, and then returns required information to the Android
application. A number of trackers, or processors as they are referred
to in BoofCV, are available. We narrowed down the choice to the
2https://developers.google.com/ar/
3https://boofcv.org/
Circulant Matrices tracker [
The Track-Learning-Detect tracker follows an object from frame to frame by localising all appearances that have been observed so far and corrects the tracker if necessary. The learning estimates the detector’s errors and updates it to avoid such errors, using a learning process. The learning process is modelled as a discrete dynamical system and the conditions under which the learning guarantees improvement are found. However, the TLD is computationally intensive. In our testing we found that when TLD was used the application would crash in older phones, due to the high memory usage.
The Circulant Matrices tracker is fast local moving-objects tracker. It uses the theory of Circulant matrices, Discrete Fourier Transform (DCF), and linear classifiers to track a target and learn its changes in appearance. The target is assumed to be rectangular with a fixed size. A dense local search, using DCF, is performed around the most recent target location. Texture information is used for feature extraction and object description. However, as only one description of the target is saved, the tracker has a low computational cost and memory footprint. Our informal in-lab testing shown that the Circulant tracker provides robust tracking.
The default implementation of the Circulant Matrices tracker in BoofCV does not work with coloured frames. Our solution first converts the video feed, one frame at a time, to greyscale using a simple RGB averaging function. The tracker returns back only the coordinates of the centre of the ROI, while the original coloured frame is rendered to the screen. Finally, to increase tracking performance, the camera is set to record at a constant resolution of 640 by 480 pixels. 2.4.5 User Interface. ABOD3-AR renders the plan directly next to the robot, as seen in Figure 2. A pivot connects the plan to the centre of the user-selected ROI. The PC-targeted version of ABOD3 ofers abstraction of information; the full plan is visible by default, but the user has the ability to hide information. This approach works on the large screens that laptops and desktops have. Contrary, at time of writing, phones rarely sport a screen larger than 15cm. Thus, to accommodate the smaller screen estate available on a phone, ABOD3-AR displays only high-level elements by default. Drives get their priority number annotated next to their name and are listed in ascending order. ABOD3-AR shares the same real-time transparency methodology as ABOD3; plan elements light up as they are used, with an opposing thread dimming them down over time.
Like its ‘sibling’ application, ABOD3-AR is aimed to be used by
both end users and expert roboticists. A study conducted by Subin
et al. [
Mechanical - Organic Artificial - Lifelike Inert - Interactive Dislike - Like
Unpleasant - Pleasant Unintelligent - Intelligent Bored - Interested Anxious - Relaxed
Group 1 (N = 23)
Group 2 (N = 22) p-value prefer to have low-level details. Hence, we added an option to enable display of the Server data, in string format, as received by ABOD3-AR. 3
A user study was carried out to investigate the efectiveness of ABOD3-AR. This took place at the University of Bath in an open public space. The study ran over five days. The principle hypothesis of this experiment is that observers of a robot with access to ABOD3-AR will be able to create more accurate mental models. In this section, we present our results, and discuss how ABOD3-AR provides an efective alternative to ABOD3 as a means to provide robot transparency. Moreover, we argue that our results demonstrate that the implementation of transparency with ABOD3-AR increases not only the trust towards the system, but also its likeability.
The R5 robot is placed in a small pen with a selection of objects,
e.g. a plastic duck. The participants are asked to observe the robot
and then answer our questionnaires. The participants are split in
two groups; Group 1 used the AR app and Group 2 did not use the
app. Participants are asked to observe the robot for at least three
minutes. A total of 45 participants took part in the experiment
(N = 45). The majority of users were aged 36 to 45. Each group had
same number of females and males. Although they worked regularly
with computers, most of them did not have a STEM background —
This was the main diference with participants in previous research
[
The Godspeed questionnaire by Bartneck et al. [
Individuals who had access to ABOD3-AR were more likely to perceive the robot as alive (M = 3.27, SD = 1.202) compare the ones without access to the app; t(43) = –0.692 and p = 0.01. Moreover, participants in the no-transparency condition described the robot as more stagnant (M = 3.30, SD = 0.926) compare to the ones in Group 2 (M = 414, SD = 0.710) who described the robot as Lively; t(43) = – 3.371, p = 02. Finally, in the ABOD3-AR condition, participants perceived the robot to be friendlier (M = 3.17, SE = 1.029) than participants in Group 1 (M = 3.77, SE = 0.869); ; t(43) = –2.104, p = 041. No other significant results were reported. These results are shown in Table 1. 3.2
We found a statistically significant diference ( p-value < 0.05) in three Godspeed questions: Dead/Alive, Stagnant/ Lively, and Unfriendly/ Friendly. The R5 has connecting wires and various chipsets exposed. Yet, participants with access to ABOD3-AR were more likely to describe the robot as alive, lively, and friendly. All three dimensions had mean values over the ‘neutral’ score of 3. Although not significantly higher, there was an indicatively increased attribution of the descriptors Interactive and Pleasant; again both with values over the neutral score. At first glance, these results suggest an increase of anthropomorphic — or at least biologic — characteristics. However, transparency decreased the perception of the robot being Humanoid and Organic; both characterizations having means below the neutral score.
Action selection takes place even when the robot is already
performing a lengthy action, e.g. moving, or when it may appears
‘stuck’, e.g. it is in Sleep drive to save battery. These results also
support that a sensible implementation of transparency, in line to
the principles set by Theodorou et al. [
An explanation for the high levels of Interest (3.8 mean for Group 1 and 4.19 mean for Group 2) is that embodied agents — unlike virtual agents— are not widely available. Participants in both groups may have bezen intrigued by the ideal of encountering a real robot. Nonetheless, our findings indicate that transparency does not necessary reduces the utility or ‘likeability’ of a system. Instead, the use of a transparency display can increase the utility and likeability of a system.
There are several characteristics of augmented reality that makes it a promising platform to provide transparency information for both industrial and domestic robots. These include the afordability of AR enabled devices, its availability on multiple platforms such as mobile phones and tablets, the rapidly increasing progress in mobile processors and cameras, and the convenience of not requiring headsets or other paraphernalia unlike its competitor virtual reality. 4
In this paper we presented a new tool, ABOD3-AR, which runs on modern mobile phones to provide transparency-related information to end users. Our tool uses a purpose-made user interface with augmented-reality technologies to display the real-time status of any robot running the Instinct planner.
As far as we are aware this is the first use of mobile augmented reality focusing solely on increasing transparency in robots and users’ trust towards them. Previous research regarding transparency in robots relied on screen and audio output or non real-time transparency. Building upon past research, we provide an afordable, compact solution, which makes use of augmented reality.
The results from a user study presented in this paper demonstrate how ABOD3-AR can be successfully used to provide real-time transparency to end users. Our results demonstrate how naive users calibrate their mental models and alter their perception of a robot as its machine nature is made cleared. Moreover, they indicate that participants with access to ABOD3-AR has higher interest to the system; potentially increasing its utility and user engagement.
The work presented in this paper is part of a research programme to investigate the efects of transparency on the perceived expectations, trust, and utility of a system. Initially this is being explored using the non-humanoid R5 robot and later we plan to expand the study using the Pepper humanoid robot manufactured by SoftBank Robotics. We argue that humanoid appearance will always be deceptive at the implicit level. Hence, we want see how explicit understanding of the robot’s machine nature efects its perceived utility. Moreover, if transparency alters trust given to the machine by its human users.
Planned future work also aims at improving the usability of the application further. Currently, the robot-tracking mechanism requires the user to manually select an area of ROI which contains the robot. Future versions of ABOD3 - AR would skip this part and replace it with a machine learning (ML) approach. This will enable the app to detect and recognize the robot by a number of features, such as colour and shape. The app will also be enhanced to be able to retrieve the robot type and plan of execution from a database of robots.