Content uploaded by Christoforos Mavrogiannis
Author content
All content in this area was uploaded by Christoforos Mavrogiannis on Nov 03, 2019
Content may be subject to copyright.
Towards Effective Human-AI Teams: The Case of Collaborative Packing
Gilwoo Lee, Christoforos Mavrogiannis, Siddhartha S. Srinivasa
Paul G. Allen School of Computer Science & Engineering
University of Washington
3800 E Stevens Way NE
Seattle, WA 98195-2350, USA
{gilwoo, cmavro, siddh}@cs.uw.edu
Abstract
We focus on the problem of designing an artificial agent (AI),
capable of assisting a human user to complete a task. Our goal
is to guide human users towards optimal task performance
while keeping their cognitive load as low as possible. Our in-
sight is that doing so requires an understanding of human de-
cision making for the task domain at hand. In this work, we
consider the domain of collaborative packing, in which an AI
agent provides placement recommendations to a human user.
As a first step, we explore the mechanisms underlying human
packing strategies. We conducted a user study in which 100
human participants completed a series of packing tasks in a
virtual environment. We analyzed their packing strategies and
discovered spatial and temporal patterns, such as that humans
tend to place larger items at corners first. We expect that im-
buing an artificial agent with an understanding of this spa-
tiotemporal structure will enable improved assistance, which
will be reflected in the task performance and the human per-
ception of the AI. Ongoing work involves the development of
a framework that incorporates the extracted insights to pre-
dict and manipulate human decision making towards an ef-
ficient trajectory of low cognitive load and high efficiency.
A follow-up study will evaluate our framework against a set
of baselines featuring alternative strategies of assistance. Our
eventual goal is the deployment and evaluation of our frame-
work on an autonomous robotic manipulator, actively assist-
ing users on a packing task.
Introduction
We consider the general scenario in which a human and an
artificial agent (AI) are collaborating to jointly complete a
task. Depending on the domain, it is often true that the ca-
pabilities of the human and the robot may greatly differ. The
AI agent may often possess superior computation capabili-
ties, whereas the human agent may have superior perceptual
and control abilities. An emerging area of research looks at
the development of frameworks that would enable effective
combined performance by leveraging the strengths of both
parties, while ensuring human comfort. Our insight is that
to achieve this goal, the AI needs to reason about the deci-
sion making of its human counterpart. At times, it may need
Copyright c
2019, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
Figure 1: We study human packing strategies. In our study inter-
face, the user sequentially drags and drops a set of objects (shown
on the right side) into a container (depicted on the left side). When
done, the user may proceed to the next task by clicking the ”done”
button (bottom left). The user is also given the options of resetting
the task to its initial state (”Reset” button), and proceeding without
completing the task (right button).
to intervene to guide their behavior towards efficiency. In
many cases, this intervention may not be realized physically
due to hardware or software limitations. In such cases, it is
critical that the AI conveys its intentions implicitly through
its actions.
In this work, we consider a scenario of collaborative pack-
ing, in which a human places a set of objects in a container
under the assistance of a recommender system. This applica-
tion is of particular relevance these days, given the increas-
ing presence of AI systems in logistics (Dekhne et al. 2019).
Achieving adequate spatial efficiency in packing is an anec-
dotally hard problem for non-expert humans, whereas ma-
nipulation is still a big challenge for robots. On the other
hand, AI systems feature superior long-term planning ca-
pabilities, whereas humans are equipped with unparalleled
manipulation capabilities. An effectively combined collab-
orative effort, leveraging the strengths of both could help
achieve increased overall performance.
As a first step to approach the outlined vision, we seek to
understand the domain of packing, focusing on the strategies
that humans employ when faced with completing a packing
task. To do so, we conducted an online user study in which
we asked human subjects to complete a series of packing
tasks in a virtual environment. Each task involved the place-
ment of a different set of 2-dimensional objects inside a
packing container. Our findings suggest that human packing
strategies in this domain can largely be classified into a set of
distinct categories corresponding to different spatiotemporal
patterns of placement. We discuss our findings and the ongo-
ing development of a planning-under-uncertainty framework
targeted towards ensuring improved efficiency and low cog-
nitive load for humans in collaborative packing scenarios.
Related work
The concept of human-AI teaming is gaining popularity, as
combining the strengths of humans and AI systems opens
promising avenues for a variety of fields and applications
(Kamar, Hacker, and Horvitz 2012; Lasecki et al. 2012;
Bayati et al. 2014; Kamar 2016). Naturally, the problem
of enabling seamless, natural, and efficient collaboration in
human-AI teams has received considerable attention over
the recent years, with researchers focusing on different as-
pects of the interaction, such as the powerful communicative
impact of actions performed in a shared context (Liang et al.
2019) or the tradeoffs between performance gains and com-
patibility with existing human mental models (Bansal et al.
2019).
For a series of applications, transferring the benefits of
human-AI teaming in the physical environment implies em-
bodiment in robot platforms. Human-robot teaming has a
unique potential for a variety of applications, given that
robots can be both intelligent and physically capable. Thus
the combination of their capabilities with those of humans
may result in performance standards that neither party could
otherwise achieve in isolation (Hoffman and Breazeal 2004).
Notable examples include fast task completion in sequential
manipulation tasks (Hayes and Scassellati 2015), and im-
proved performance through intelligent resource allocation
to human participants (Jung et al. 2018).
A common complication in such applications is that ex-
plicit communication between the human and the robot is
often not feasible, effective or desired. Therefore, in or-
der to be of assistance, the robot needs to infer the inten-
tions of the user implicitly, through observation of their ac-
tions, and clearly communicate its own intentions, through
its own actions (Knepper et al. 2017). A typical paradigm
of particular relevance in this domain is shared autonomy,
in which a robot assists a human user in completing their
task. In a variety of applications, it has been shown that in-
ferring and adapting to human intentions is positively per-
ceived by users and effective (Dragan and Srinivasa 2013;
Kuderer et al. 2014; Gopinath, Jain, and Argall 2017; Jav-
dani et al. 2018). Furthermore, understanding the mecha-
nisms underlying human decision making in a particular do-
main is shown to yield performance improvements and pos-
itive impressions in joint tasks (Nikolaidis and Shah 2013;
Nikolaidis et al. 2015). Finally, explicitly collaborative tasks
such as collaborative manipulation (Dragan and Srinivasa
2014; Nikolaidis, Hsu, and Srinivasa 2017), and assembly
(Knepper et al. 2015) or implicitly collaborative tasks such
as social navigation (Mavrogiannis et al. 2019) benefit sig-
nificantly by the incorporation of models of human infer-
ence.
In this work, we consider a joint task (packing), per-
formed in collaboration between a human and an AI agent.
We also consider a setting of implicit communication, in the
sense that human intentions are not directly observed and
need to be inferred. Our first step towards approaching this
scenario is to understand the domain by collecting and ana-
lyzing human data.
Study design
Our study was conducted online on an interactive web appli-
cation. Participants were recruited online, through the Ama-
zon Mechanical Turk platform (Buhrmester, Kwang, and
Gosling 2011). Upon providing consent, each participant
was assigned the same set of 65 packing tasks, presented in a
random order. These tasks involved the placement of sets of
4-8 rectangular objects of different sizes inside a rectangular
container of fixed size.
Interface
The web interface depicts a set of rectangular objects of var-
ious dimensions, alongside a rectangular container, from a
top view (see Fig. 1). Participants were instructed to sequen-
tially place all of these objects at locations of their choice,
inside the container. Once an object is placed inside the con-
tainer, it cannot be moved – thus participants are forced to
judiciously decide on the placements of their objects. The in-
terface comprises three buttons: (a) a button for proceeding
to the next task (shown at the bottom left); (b) a reset button
(middle), useful in cases where participants’ decisions did
not allow them to put all object to the container; (c) a button
which allowed participants to proceed to the next task with-
out completing the current one (right). Since at this stage
we were interested in understanding the domain of packing
rather than participants’ performance, we gave users the op-
tion of resetting a task to its initial state by hitting the “Re-
set” button. We also gave participants the option to skip a
task if they decided to but we disincentivized this option by
placing a sad face on the corresponding button. Similarly,
we incentivized completion by placing a happy face on the
“Done” button.
Generation of Packing Tasks
The complexity of a packing task depends on the relative
sizes of objects among themselves and with respect to the
container. In practice, these relationships yield different tol-
erance requirements in the object placements. The smaller
the tolerances, the higher the amount of precision required
to ensure a collision-free placement, and thus the more com-
plex the packing task becomes.
While we can generate arbitrarily easy or complex pack-
ing tasks by designing a large container and many small
(a) Example packing instance (b) Spatial clusters
AB AC AD AE BA CA
0
20
40
Frequency (%)
(c) Temporal patterns of first 2 items
Figure 2: Highlights from our study. We found a few strong spatiotemporal patterns in people’s packing styles, such as placing large items
into corners or packing larger items before smaller ones. (a) An example packing instance by a user. This person packed in the order of A-C-
B-D-E. (b) Spatial clusters of configurations for the items in (a). Because people tend to put larger items into corners, the final configurations
can be clustered into a few spatial patterns. For this task, 4 strong spatial patterns are shown. (c) Temporal patterns of the first two items for
this task. Nearly everyone chose the largest item (A) first, and 85% of them picked one of the two second largest items (B or C).
items that could be packed in many different ways, such
problem instances would not help us observe any discernible
patterns that humans may naturally have. In order to iden-
tify spatiotemporal patterns in packing, we have designed
our packing tasks such that each task satisfies the following
conditions:
•At least 70% of the container is filled with the items.
•There is a finite number of clusters of spatially feasible
solutions.
By committing to these conditions, we constrain the tasks
to be doable with a finite number of qualitatively equivalent
object placements. Our expectation was that even under the
constrained setting of finite spatially feasible solutions, in-
nate human packing styles and preferences would still mani-
fest themselves. In particular, we expected that human pack-
ing strategies would show strong inclinations towards dis-
tinct classes of spatiotemporal placements.
In order to generate packing tasks that satisfy the above
conditions, we fix the size of the container, randomly gen-
erate items of various sizes, and then attempt to place them.
For any resulting placement, if more than 70% of the con-
tainer is filled with 4-8 items, then we test if the second con-
dition is met. To do so, we empty the container and attempt
to place the same items in different ways. If we can generate
more than 50 different collision-free configurations, then we
run a Principal Components Analysis (PCA) (Jolliffe 1986)
on the configurations. Each configuration is represented with
a vector x∈R2n, stacking the Cartesian coordinates of all n
items. We take the first two dimensions of the PCA projec-
tion, visually check for discernible clusters as in Fig. 2(b),
and keep the task only if such clusters are found. In total,
we have generated 65 tasks of varying complexity with the
following distribution: 17 tasks of 4 objects, 15 tasks of 5
objects, 20 tasks of 7 objects, and 10 tasks of 8 objects.
Dataset & Analysis
We had 100 participants (34 female, 66 male), recruited
through the Amazon Mechanical Turk platform. The par-
ticipants were between 18 and 65 years old (M= 32.08,
SD = 8.87). Each participant was given the 65 tasks in a
randomly generated order and was asked to complete them
within an hour. Although some participants did not com-
plete all tasks, all of them were completed by roughly the
same number of participants, and the distribution was the
following: 4-objects tasks (M= 89.59,SD = 1.94); 5-
object tasks (M= 85.60,SD = 3.07); 6-object tasks
(M= 79.00,SD = 1.41); 7-object tasks (M= 79.10,
SD = 3.58); 8-object tasks (M= 78.20,S D = 1.54).
On average, participants took about 40 minutes to com-
plete the tasks. For each task, we recorded the ordering and
object placement locations inside the container. The col-
lected dataset is grouped per task class (a task class is a set
of tasks with the same number of objects). For each task
class: (a) we cluster the recorded solutions with respect to
their spatial patterns using Principal Components Analysis
(PCA) (see); (b) we classify the provided solutions into a
set of classes by looking at the first two item placements
and comparing the frequency of each ordered pair. Fig. 2
depicts an example packing task completed by a participant,
and illustrates the associated packing trends extracted with
the outlined process.. Fig. 2(b) illustrates four distinct spa-
tial clusters of object placements that emerged in subjects’
placements. For the same task, Fig. 2(c) describes the fre-
quency of different temporal patterns that emerged.
To extract a more holistic view of the data, we look at
the distributions of placement and ordering strategies per
task class. Specifically, for each task class, we count the fre-
quency that each placement cluster occurred, and the fre-
quency that each ordering occurred. To quantify the relative
preferences of participants over different strategies, we com-
pute the information entropy for each the frequency distribu-
tions. Intuitively, entropy quantifies how uniform the distri-
bution was (higher entropy indicates higher uniformity in the
distribution). To make sure the entropy calculations are re-
latable across task classes, we normalize the raw values with
the maximum theoretical entropy per task class (the maxi-
mum entropy corresponds to the event that all clusters are
Figure 3: Normalized Information Entropy of placement frequency
per task. The Yellow lines indicate mean entropy per task class,
whereas error bars correspond to standard deviations. The entropy
stays high across the tasks, indicating that the participants did not
have strong preference among the feasible spacial clusters.
equally represented in a task class).
Fig. 3 depicts the normalized entropy of the placement
frequency distribution, whereas Fig. 4 shows the normalized
entropy of the ordering frequency distribution. We see rela-
tively high entropy in the placement strategies across all task
classes. High entropy in this case indicates that the place-
ment clusters tended to be roughly equally represented in
participant’s packing strategies. Although placement trends
seem to exist, they are not as strong as we expected. Further
visual inspection of the examples showed that participants
tended to place larger items at the corners of the container.
This could be the result of human preferences and decision
making, but it could also be an artifact of the constrained de-
sign. Regarding the ordering, we see relatively low-entropy
frequencies across all task classes. This indicates the ex-
istence of a temporal structure in subjects’ strategies, i.e.,
participants were partial towards specific orderings. Qual-
itatively, we have observed that participants tend to choose
larger objects first, put them closer to one of the four corners,
and then pack smaller items in descending order in size.
Discussion
The findings of this study illustrate our extracted knowl-
edge about the particular domain in consideration, i.e., that
of 2-dimensional packing. We discovered that human pack-
ing strategies in this domain tend to follow specific spa-
tiotemporal patterns. These patterns are indicative of spa-
tial and temporal preferences of humans in packing tasks.
It appeared that the temporal preferences were quite strong
(see Fig. 4), whereas the spatial preferences not as much
(see Fig. 3). Additional qualitative examination of the dis-
covered patterns provided insights into the types of these
preferences. Notable examples include participants’ inclina-
tion towards placing larger objects in the beginning of the
task, and placing larger objects at the corners of the con-
Figure 4: Normalized Information Entropy of order frequency
per task. The Yellow lines indicate mean entropy per task class,
whereas error bars correspond to standard deviations. Even as the
number of objects per task increases, the entropy stays low, indi-
cating strong temporal patterns.
tainer. We expect that identifying and adapting to observed
packing strategies online could enable an artificial agent to
assist a human agent effectively.
Ongoing & Planned Work
Our key insight is that understanding the mechanisms un-
derlying human decision making could enable an artificial
agent to provide effective assistance, yielding improved task
performance and reducing cognitive load for human users.
Some domains can be particularly challenging for humans,
for reasons related to the limits of human computational
abilities. For example, in the packing domain, the limited
human planning horizon and human spatial efficiency can
greatly affect task performance and mentally load humans
to an undesired extent. In fact, packing can be cast as the
knapsack optimization problem, which is known to be NP-
hard (Garey and Johnson 2002). We expect that an AI agent,
capable of modeling both the knapsack problem and the
mechanisms underlying human decision making could pro-
vide effective assistance resulting in improved task perfor-
mance and reduced cognitive load for human users. Ongo-
ing work involves the development of a planning framework
that would allow us to test this hypothesis through a follow-
up user study.
A Framework for Packing Assistance
Motivated by the findings of the presented study, we develop
a framework for planning under uncertainty that incorpo-
rates modeling of human decision making in collaborative
packing tasks. In particular, we are working on adapting
the Bayesian Reinforcement Learning (BRL) framework of
Lee et al. (2019) to enable reasoning about uncertainty over
human packing strategies. BRL is a reinforcement learn-
ing framework that incorporates a mechanism for reasoning
about model uncertainty. It models the problem as a Bayes-
Adaptive Markov Decision Process (BAMDP) (Duff 2002),
explicitly modeling uncertainty as a belief over a latent un-
certainty variable, incorporated in the transition function and
the reward function. Overall, BRL maximizes the expected
discounted reward, given the uncertainty. We believe that
this mechanism is of particular relevance and value in prob-
lems involving human interaction, where uncertainty is typ-
ically over human mental models underlying their decision
making.
For our task domain, we are incorporating a belief dis-
tribution over the human user’s spatiotemporal placement
strategy, given the container configuration and the object’s
shape. We plan on using the collected human dataset to learn
the outlined predictive model. During execution, we will be
using our framework as a recommender system that will be
providing online recommendations to the human user.
Planned User Study
To formally investigate our outlined insight, we design an
online user study, in which human subjects will be exposed
to a set of conditions (within-subjects), corresponding to dif-
ferent modes of AI assistance. More specifically, we con-
sider the following set of conditions:
1. No recommendation – the user completes the task without
receiving any assistance.
2. The system provides object recommendations, i.e., assists
by manipulating the order of object placements.
3. The system provides both order and placement recom-
mendations.
4. The system provides random object recommendations.
5. The system provides random order and random placement
recommendations.
We hypothesize that the assistive conditions will yield
improved task performance compared to the condition of
no assistance, but also more positive human ratings and re-
duced reported cognitive load. As performance metrics, we
consider the time-to-completion and the packing spatial ef-
ficiency. After each condition, we will collect ratings of
perceived system intelligence, likeability, and predictability,
based on the Godspeed (Bartneck, Croft, and Kulic 2009)
to understand the perception of the considered conditions
from the perspective of participants. Finally, we will mea-
sure the cognitive load associated with each condition by
presenting a questionnaire based on the NASA-TLX (Hart
and Staveland 1988). Finally, participants will be provided
with an open-form question, asking them to provide qualita-
tive feedback of their choice regarding their interaction with
the system.
Acknowledgments
Gilwoo Lee is partially supported by Kwanjeong Educa-
tional Foundation. This work was partially funded by the
Honda Research Institute USA, the National Science Foun-
dation NRI (award IIS-1748582), and the Robotics Collab-
orative Technology Alliance (RCTA) of the United States
Army Laboratory.
References
Bansal, G.; Nushi, B.; Kamar, E.; Weld, D.; Lasecki, W.; and
Horvitz, E. 2019. Updates in human-ai teams: Understand-
ing and addressing the performance/compatibility tradeoff.
In Proceedings of the AAAI Conference on Artificial Intelli-
gence.
Bartneck, C.; Croft, E.; and Kulic, D. 2009. Measurement
instruments for the anthropomorphism, animacy, likeability,
perceived intelligence, and perceived safety of robots. Inter-
national Journal of Social Robotics 1(1):71–81.
Bayati, M.; Braverman, M.; Gillam, M.; Mack, K. M.; Ruiz,
G.; Smith, M. S.; and Horvitz, E. 2014. Data-driven de-
cisions for reducing readmissions for heart failure: General
methodology and case study. PLOS ONE 9(10):1–9.
Buhrmester, M.; Kwang, T.; and Gosling, S. D. 2011. Ama-
zon’s mechanical turk: A new source of inexpensive, yet
high-quality, data? Perspectives on Psychological Science
6(1):3–5.
Dekhne, A.; Hastings, G.; Murnane, J.; and Neuhaus, F.
2019. Automation in logistics: Big opportunity, bigger un-
certainty. McKinsey & Company. Retrieved: 2019-10-31.
Dragan, A. D., and Srinivasa, S. S. 2013. A policy-blending
formalism for shared control. The International Journal of
Robotics Research 32(7):790–805.
Dragan, A. D., and Srinivasa, S. 2014. Integrating hu-
man observer inferences into robot motion planning. Au-
tonomous Robots 37(4):351–368.
Duff, M. O. 2002. Optimal Learning: Computational Proce-
dures for Bayes-adaptive Markov Decision Processes. Ph.D.
Dissertation.
Garey, M. R., and Johnson, D. S. 2002. Computers and
intractability, volume 29. wh freeman New York.
Gopinath, D.; Jain, S.; and Argall, B. D. 2017. Human-
in-the-loop optimization of shared autonomy in assistive
robotics. IEEE Robotics and Automation Letters 2(1):247–
254.
Hart, S. G., and Staveland, L. E. 1988. Development of nasa-
tlx (task load index): Results of empirical and theoretical
research. In Hancock, P. A., and Meshkati, N., eds., Human
Mental Workload, volume 52 of Advances in Psychology.
North-Holland. 139 – 183.
Hayes, B., and Scassellati, B. 2015. Effective robot
teammate behaviors for supporting sequential manipulation
tasks. In Proceedings of the IEEE/RSJ International Confer-
ence on Intelligent Robots and Systems (IROS), 6374–6380.
Hoffman, G., and Breazeal, C. 2004. Collaboration in
human-robot teams. In Proceedings of the AIAA Intelligent
Systems Technical Conference, 6434.
Javdani, S.; Admoni, H.; Pellegrinelli, S.; Srinivasa, S. S.;
and Bagnell, J. A. 2018. Shared autonomy via hindsight op-
timization for teleoperation and teaming. The International
Journal of Robotics Research 37(7):717–742.
Jolliffe, I. 1986. Principal Component Analysis. Springer
Verlag.
Jung, M. F.; DiFranzo, D.; Stoll, B.; Shen, S.; Lawrence, A.;
and Claure, H. 2018. Robot assisted tower construction-a
resource distribution task to study human-robot collabora-
tion and interaction with groups of people. arXiv preprint
arXiv:1812.09548.
Kamar, E.; Hacker, S.; and Horvitz, E. 2012. Combining
human and machine intelligence in large-scale crowdsourc-
ing. In Proceedings of the International Conference on Au-
tonomous Agents and Multiagent Systems (AAMAS), 467–
474.
Kamar, E. 2016. Directions in hybrid intelligence: Com-
plementing ai systems with human intelligence. In Proceed-
ings of the International Joint Conference on Artificial Intel-
ligence (IJCAI), 4070–4073.
Knepper, R. A.; Tellex, S.; Li, A.; Roy, N.; and Rus, D. 2015.
Recovering from failure by asking for help. Autonomous
Robots 39(3):347–362.
Knepper, R. A.; Mavrogiannis, C. I.; Proft, J.; and Liang, C.
2017. Implicit communication in a joint action. In Proceed-
ings of the ACM/IEEE International Conference on Human-
Robot Interaction (HRI), 283–292.
Kuderer, M.; Sprunk, C.; Kretzschmar, H.; and Burgard, W.
2014. Online generation of homotopically distinct naviga-
tion paths. In Proceedings of the IEEE International Con-
ference on Robotics and Automation (ICRA), 6462–6467.
Lasecki, W. S.; Bigham, J. P.; Allen, J. F.; and Ferguson,
G. 2012. Real-time collaborative planning with the crowd.
In Proceedings of the AAAI Conference on Artificial Intelli-
gence, 2435–2436.
Lee, G.; Hou, B.; Mandalika, A.; Lee, J.; and Srinivasa, S. S.
2019. Bayesian policy optimization for model uncertainty.
In Proceedings of the International Conference on Learning
Representations (ICLR).
Liang, C.; Proft, J.; Andersen, E.; and Knepper, R. A.
2019. Implicit communication of actionable information in
human-ai teams. In Proceedings of the CHI Conference on
Human Factors in Computing Systems, 95:1–95:13.
Mavrogiannis, C.; Hutchinson, A. M.; Macdonald, J.; Alves-
Oliveira, P.; and Knepper, R. A. 2019. Effects of distinct
robot navigation strategies on human behavior in a crowded
environment. In Proceedings of the ACM/IEEE Interna-
tional Conference on Human-Robot Interaction (HRI), 421–
430.
Nikolaidis, S., and Shah, J. 2013. Human-robot cross-
training: Computational formulation, modeling and evalu-
ation of a human team training strategy. In Proceedings of
the ACM/IEEE International Conference on Human-robot
Interaction (HRI), 33–40.
Nikolaidis, S.; Ramakrishnan, R.; Gu, K.; and Shah, J. 2015.
Efficient model learning from joint-action demonstrations
for human-robot collaborative tasks. In Proceedings of the
ACM/IEEE International Conference on Human-Robot In-
teraction (HRI), 189–196.
Nikolaidis, S.; Hsu, D.; and Srinivasa, S. 2017. Human-
robot mutual adaptation in collaborative tasks: Models and
experiments. The International Journal of Robotics Re-
search 36(5-7):618–634.
