Academics involved in the study from Imperial College London, France and Japan have previously shown that where two people work together on a physical task, this improves the performance of both participants, for example, while performing a paired dance like the Tango.
To test this they had two human partners move joystick-like devices connected by a virtual elastic band to follow targets on a computer screen (see video) and showed that non-verbal clues from the partner helped each person’s performance.
In the new study, published today in Nature Human Behaviour, the same group of academics used robots to explore how non-verbal cues help to improve performance. They suggested that a human participant will use their sense of touch and proprioception (their sense of position of self and movement) to estimate and predict their partner’s next action and be able to match them, causing both to improve continuously throughout the task.
To test their theory, they paired human participants with robots that were programmed to perform either better or worse at the task than the human’s ability. They used the same joystick task as the aforementioned study to test the theory.
They found that the robots that were ‘better’ than their humans appeared to correct, or compensate for, the human’s movements to make them perform better. The robots that were ‘worse’ than their humans also somehow improved the performance of their human partners.
Lead author Atsushi Takagi, PhD student from Imperial’s Department of Bioengineering, said: “This is the way a human partner would learn to work with a fellow human, as one would get used to the other’s working patterns and become able to predict their moves.”
Overall, when using programmed robots as partners, humans performed just as well on the task as they did when paired with other humans.
The human subjects performed worst of all when they attempted the task with no human or robot partner, suggesting, like the previous study, that people learn to complete physical tasks to much higher standards when interacting with a partner.
The authors used Bayesian optimisation, a statistical technique that forms the basis of machine learning, to allow the robot to ‘learn’ the behaviour of the human partner from previous movements and optimally assist their movement.
For example, the robot would quickly learn that particular on-screen movements by the target would cause their particular human to overshoot at the target. In this case, the robot would compensate by undershooting for that same target, hopefully meeting the human’s efforts somewhere in the middle and coming to a more accurate result.
Takagi added: “Although only demonstrated with joysticks and computer programmes so far, we feel that in future, there may be such a thing as a ‘robotic physiotherapist’. Nobody wants to replace a human physiotherapist, but the number of hours they can spend with patients is limited. A robot physiotherapist could complement a human one, taking over when they have to move on to the next patient. Ultimately, this may speed up recovery times for patients.”
Co-author Professor Etienne Burdet, from Imperial’s Department of Bioengineering, added: “Robotic therapists could also act as an extension of their human counterparts. I could imagine a scenario where a number of robots, carrying out treatments with patients in the comfort of their homes, are supervised remotely by a physiotherapist in a hospital location. This could magnify the efforts of a physiotherapist and help more people to get the therapies they need.”