Science Robotics | Focus
HUMAN-ROBOT INTERACTION Copyright © 2024 The
Authors, some rights
Crucial hurdles to achieving human-robot harmony reserved; exclusive
licensee American
Association for the
Keya Ghonasgi1, Taylor Higgins2*, Meghan E. Huber3, Marcia K. O’Malley4 Advancement of
Science. No claim to
Holistic consideration of the human and the robot is necessary to overcome hurdles in human-robot interaction. original U.S.
Government Works
A hallmark of humans is our ability to create only through more careful study of the hu- perception is inversely related to interaction
tools that enhance our capabilities, and in- man component. complexity given that the human sensing
telligent robots are the most sophisticated The effect of physical interaction with a block in Fig. 1 is likely to be imperfect and
and powerful of such tools to date. Robots robot on human behavior can be formally inconsistent across individuals. A major
Downloaded from https://www.science.org at Georgia Institute of Technology on November 14, 2024
can enhance human motor function, learn studied through the natural versus forced challenge with modeling human sensing
to refine and elevate their own autonomous framing. Figure 1 illustrates this concept, lies in our limited ability to measure how
skills, and offer the potential to go be- where the two signals entering the human the human senses the robot and the task.
yond inherent human and robot abilities feedback loop are the task outcomes (yellow As a result, the process through which the
through collaboration. Realizing the ben- signal) and the output of the robotic plant human uses information from various sens-
efits of such shared autonomy, however, is block (orange signal). The observed human ing modalities (such as visual, propriocep-
not guaranteed. Natural human adaptation behavior (purple dashed signal) in the ab- tive, tactile, and auditory modalities) to
and programmed robot adaptation may sence of robot interaction would represent interpret the actions and intentions of a
not synergize effectively, leading to collab- the human system’s natural response to the robot partner remains elusive. Psychological
orative task performance that is no better, task. The additional interaction with the ro- and neurological approaches to studying so-
or even worse, than if either agent per- bot, represented as the solid orange line cial human-robot interactions have recently
formed the task alone. Moreover, subopti- from the robot plant to the summation be- shown some promise toward addressing
mal human-robot interaction in the physical fore the human plant in Fig. 1, changes hu- this gap (4).
world can have severe consequences, such as man behavior. The observed behavior is The human control block oversees high-
human injury and equipment damage. Thus, then a sum of the natural response (yellow level decisions and their enactment by the
prudent design of collaborative robot algo- input) to the task and the forced response to central nervous system in response to envi-
rithms is essential to prevent negative con- the robot (orange input). ronmental inputs (5). This dynamic re-
sequences stemming from ineffective and Effective robot adaptation necessitates an sponse, influenced by decision-making,
unsafe interactions with humans. adequate model of the natural or the forced adaptation, and motor training, as well as
Collaboration between humans and ro- human feedback loop response, depending cognitive factors like expectation, trust, and
bots is heavily influenced by the bidirection- on the specific task. Some pHRI paradigms, attention levels, introduce variability, a key
al nature of the interaction between two (or such as human-in-t he-loop optimization aspect of human movement fostering adapt-
possibly more) agents, as demonstrated in (1), only model the behavior of the coupled ability (6). However, such variability also
the block diagram depicted in Fig. 1. The human-robot system (forced response). Al- makes predicting human behavior difficult
human and robot feedback loops show each ternatively, in applications such as robot- for a robot.
agent’s control (decision-making), plant aided rehabilitation, isolating the natural Despite its adaptability, researchers of-
(physical embodiment), and sensing (per- response of the human is necessary. In mo- ten characterize the human control block
ception) blocks. Interaction arrows link the tor control and learning studies, the human as a simplified and time-invariant system,
robot, human, and task blocks, symbolizing natural response is tested by randomly turn- ignoring the potential for humans to learn
the simultaneous and bidirectional ex- ing off the robot interaction during “catch to interact with robots. In addition, hu-
change of actions and real-time responses trials” (2). Methods that similarly evaluate mans are often assumed to be inherently
between humans and robots toward a com- and monitor the human response to robot optimal. Although some work has shown
mon task goal. Although achieving optimal interaction, especially over long timescales, the optimality of human behavior in specific
shared performance demands consideration need more attention to quantify pHRI settings (7) and the human motor control
of both feedback loops, the robot loop has success. system has been likened to adaptive control
received far more attention than the human The success of human-robot interaction laws that we use in robotics (8), a unified cost
loop in the literature. Here, we contend that depends on each agent’s ability to accurately function for all behavior remains elusive.
the next major breakthrough in physical sense and perceive the interaction and the Many sources of complexity—the intractable
human-robot interaction (pHRI) will occur collaborative goal (3). The accuracy of this number of influencing variables, the individu-
alized nature of human behavior and learn-
1
George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA 30332, ing, and the complications introduced by a
USA. 2Department of Mechanical Engineering, Florida A&M University–Florida State University, Tallahassee, FL robotic agent when the human has little
32310, USA. 3Department of Mechanical and Industrial Engineering, University of Massachusetts Amherst,
Amherst, MA 01003, USA. 4Department of Mechanical Engineering, Rice University, Houston, TX 77005, USA. prior experience and trust—make it diffi-
*Corresponding author. Email: thiggins@eng.famu.fsu.edu cult to find a universal human control
Ghonasgi et al., Sci. Robot. 9, eadp2507 (2024) 13 November 2024 1 of 2
Science Robotics | Focus
The goal is to create useful models that ad-
equately describe the aspects of the human
feedback loop that act as a bottleneck to
achieving the promise of shared human-
robot interactions.
REFERENCES
1. P. Slade, C. Atkeson, J. M. Donelan, H. Houdijk,
K. A. Ingraham, M. Kim, K. Kong, K. L. Poggensee,
R. Reiner, M. Steinert, J. Zhang, S. H. Collins, On
human-in-the-loop optimization of human-robot
interaction. Nature 633, 779–788 (2024).
2. J. Kluzik, J. Diedrichsen, R. Shadmehr, A. J. Bastian,
Reach adaptation: What determines whether we learn
an internal model of the tool or adapt the model of our
arm? J. Neurophysiol. 100, 1455–1464 (2008).
Downloaded from https://www.science.org at Georgia Institute of Technology on November 14, 2024
3. J. Lee, M. E. Huber, N. Hogan, Gait entrainment to torque
pulses from a hip exoskeleton robot. IEEE Trans. Neural
Fig. 1. Schematic of the bidirectional human-robot system. This schematic represents the interactions Syst. Rehabil. Eng. 30, 656–667 (2022).
4. E. S. Cross, R. Hortensius, A. Wykowska, From social
between the internal feedback loops of the human and the robot and highlights the considerations necessary
brains to social robots: Applying neurocognitive
to perform a common task.
insights to human–robot interaction. Philos. Trans. R.
Soc. B 374, 20180024 (2019).
5. C. Fang, L. Peternel, A. Seth, M. Sartori, K. Mombaur,
model (9). Future work should address each Creating models of the human perception- E. Yoshida, Human modeling in physical human-robot
area of complexity both separately and cognition-action system is critical for en- interaction: A brief survey. IEEE Robot. Autom. Lett. 8,
holistically. abling new, successful behaviors to emerge 5799–5806 (2023).
The human neuromotor system is pro- synergistically from the physical interac- 6. N. Stergiou, L. M. Decker, Human movement variability,
nonlinear dynamics, and pathology: Is there a
foundly complex and should not be under- tions between humans and robots. However,
connection? Hum. Mov. Sci. 30, 869–888 (2011).
estimated when developing responsive building precise models of the entire human 7. J. C. Selinger, J. D. Wong, S. N. Simha, J. M. Donelan, How
robotic interactions. First-principles ap- neuromotor system is, at this time, unrealis- humans initiate energy optimization and converge on their
proaches offer analytical insights but are tic and impractical. For centuries, neuro- optimal gaits. J. Exp. Biol. 222, jeb198234 (2019).
constrained by oversimplified assumptions scientists have sought to mathematically 8. E. Todorov, Optimality principles in sensorimotor
control. Nat. Neurosci. 7, 907–915 (2004).
about model structure. Conversely, data- describe human sensing, control, and move- 9. E. C. Yu, D. A. Lagnado, The influence of initial beliefs on
driven black-box methods are limited by ment execution. Although we urge the pHRI judgments of probability. Front. Psychol. 3, 381 (2012).
training data quality and quantity. This research community to move beyond ignor- 10. H. L. Fernandes, K. P. Kording, In praise of “false”
manuscript aims to call the community to ing or oversimplifying the human feedback models and rich data. J. Mot. Behav. 42, 343–349
(2010).
action to leverage a judicious combination loop, we also challenge the community to
of these approaches toward empirically focus on developing models that delicately
modeling human-robot interactive behaviors. balance complexity and competence (10). 10.1126/scirobotics.adp2507
Ghonasgi et al., Sci. Robot. 9, eadp2507 (2024) 13 November 2024 2 of 2
