Devices and experimental setups
All studies included in this review were based on upper-limb movements or rehabilitation tasks. Most studies were also performed with hand-held end-effector-type manipulators, haptic devices, or sensors where there is a single interaction point between the device and the user. These studies could be extended to lower-limb movements (e.g., tracking trajectories with the ankle for rehabilitation [48]) to investigate if current results can be generalized. Furthermore, conducting studies with multiple interaction points between the device and the user, such as with a multiple DOF exoskeleton that can provide virtual connections of varying and controllable impedance at the joint level between dyads, would aid our understanding of how dyadic interaction affects task performance and motor learning at the joint level.
Haptic rendering
In most studies on HRH physical interaction, the haptic connections between dyads are virtual spring-damper elements. Other physical interaction schemes, such as magnet-like force fields [21], could be investigated to provide a holistic view of the effects of physical interaction. Lessons can also be learned from the vast amount of previous work in teleoperation [49,50,51,52]. A promising possible approach, due to its similarity to traditional rehabilitation, could be to mirror the motion of user A to user B, while only sending the force measurement of user B to user A [53,54,55,56]. This type of asymmetric physical interaction can be beneficial, especially in cooperative scenarios where users have different roles. Designing the characteristics of the physical interaction based on the assigned roles could improve the outcomes of the interaction. For example, in a rehabilitation setting, the therapist can follow the patient’s motion through their own device while applying forces as needed to assist or correct the patient instead of just being connected with virtual springs [14].
Towards haptic rendering in a network of subjects
While most studies focus on interaction between only two subjects, scenarios where a network of people interact with each other can be quite beneficial, especially in a rehabilitation setting. For example, the ability to simultaneously interact with a network of patients can allow therapists to provide more efficient and effective therapy. Even though it might be expected that adaptation and learning will take more time with more people, especially if the interaction is only non-physical, Takagi et al. [12] obtained promising results related to the benefits of increased group size with physical interaction. They examined the task performance of dyads, triads and tetrads whose hands are connected with virtual springs while they follow a common moving target. They found that individuals use the interaction force to estimate the collective’s target and improve their movement planning. As the group size increases, the variance of haptic inference decreases, which results in better estimate and task performance.
Considering the initial positive results and its possible impact on rehabilitation practice, we believe that it will be beneficial to the field to allow a network of people to interact with each other through their own robotic devices in future studies. In addition to solving the practical challenges of developing these systems, the effect of multilateral HRH interactions on task performance and motor learning should be explored.
Metrics
Task-related measurements (e.g., completion time, tracking error) are generally used to quantify short-term motor learning due to dyadic training. Motor learning also involves neuroplastic changes in the brain, and it has two key attributes: retention and transfer. These are not studied in the papers we reviewed. With advanced technologies in electroencephalography (EEG), we could potentially measure the neuroplastic changes in the brain during different dyadic interaction conditions and further understand the mechanism of dyadic interactions in motor learning [57, 58]. In addition, it would be useful to include experiments that investigate how the key factors (e.g., interaction type, interaction mode, partner characteristics) in dyadic interactions affect the retention/persistence and transfer/generalization of motor learning, which is essential for rehabilitation purposes.
Two papers [3, 22] examine how the relationship between partners affects their motivation and preferences, and many studies show that motivation can directly influence training intensity, which improves functional outcomes. However, there has been little study of the direct effect of different interaction modes on clinical outcomes (e.g., via the Fugl-Meyer Assessment). Clinical measurements should be used in combination with subjective and objective measurements to assess how key factors in dyadic interaction affect rehabilitation outcomes.
Effects of the interaction type and characteristics
Twelve studies [2, 15,16,17,18,19, 28, 29, 36,37,38, 40] that analyzed the effects of interaction type and characteristics mostly reported similar results on dyadic interaction. However, there are several contradicting results and areas that suggest further investigation.
Limited studies about the effects of interaction characteristics on individual motor learning
Some tasks can only be performed by two people, so only dyadic task performance can be analyzed. In a rehabilitation setting, however, the goal is motor learning for the individual patient. In other words, in rehabilitation settings, dyadic interaction is used to train people to improve their solo performance. However, most studies investigate the effects of interaction types and characteristics focused on dyadic task performance but do not investigate motor learning of individuals due to dyadic training.
Only Beckers et al. [37] investigated the effect of interaction characteristics (i.e., spring stiffness) on individual motor learning and found no significant learning difference between training with compliant and stiff interaction. Considering the fact that they also did not find any learning difference between solo and stiff connection, it would be valuable to replicate a similar experimental procedure with much stiffer virtual springs or with rigid interaction. This would allow to compare the two extremes of the physical interaction characteristics (no interaction and rigid interaction).
Conflicting results on the effect of interaction stiffness
Four studies implemented collaborative target tracking tasks where solo and dyadic tracking performances are compared [2, 15, 28, 37]. While Ganesh et al. [2] found that the biggest dyadic task performance improvement (with respect to solo) occurs at intermediate levels of stiffness between members of a dyad, Che et al. [28] , Takagi et al. [15] and Beckers et al. [37] observed that as the stiffness increases performance improvement increases. (The stiffness values are given in Additional file 2: Table S3.) The reasons for these different results are unclear, but two main differences between the experimental setups might be the source of the conflicting results. First, the tracking tasks in the study of Takagi et al. [15] and Che et al. [28] were 1 DoF, but in the study of Ganesh et al. [2] and Beckers et al. [37], subjects tracked targets that moved in a plane. Second, while in the articles of Ganesh et al. [2], Takagi et al. [15], and Beckers et al. [37] subjects tracked a moving target, in the study of Che et al. [28] subjects reached to a static target. Therefore, additional studies with different task spaces and target types (moving, static) could be of help to better understand the effect of interaction stiffness on task performance.
Little study of the effect of interaction damping
Even though several studies [2, 28, 29, 37] implemented virtual dampers between subjects for stability, they were focused on the effect of different spring stiffnesses on dyadic task performance. Only Tanaka et al. [16] analyzed the effects of virtual damping by varying its viscosity. They found significantly different task performance at different viscosities and damping ratios. Additional study of the effect of velocity-dependent forces on task performance is warranted.
Several studies have reported promising motor learning results for single-person upper-limb reaching or tracking tasks due to training under destabilizing negative viscosity [59,60,61]. Negative viscosity can improve motor learning by increasing motor variability and facilitating the development of internal models of body dynamics [62, 63]. However, none of the references reviewed in this paper implemented negative viscosity between the subjects.
Effects of the interaction mode
Many factors contribute during non-physical interaction
It is clear that, when the interaction mode is selected properly, non-physical interaction can improve the participants’ motivation, engagement, and gaming experience, leading to an increase in physical activity and training intensity. Many of the included studies confirmed that the personality of the user, the intimacy between dyads, and environmental factors play important roles in choosing the proper interaction mode. In neuromuscular therapy, training intensity—alongside early treatment and user-centered, task-oriented training—is a key factor for functional improvement. Therefore, non-physical interaction has great potential to further increase the benefits brought by robot-assisted neuromuscular and virtual reality-assisted therapy.
Comparison of interaction modes during physical interaction is needed
Despite the increasing interest in comparing different interaction modes for non-physical interaction conditions, no study compares how different physical interaction modes affect task performance or individual motor learning. Moreover, no study has implemented competitive physical interaction between subjects among the selected studies.
Even though competitive physical interaction would result in worse task performance during the training, it is quite promising for improved individual motor learning. Considering the positive results obtained by haptic error modulation and resistive training studies [64,65,66,67], competitive interaction can facilitate error-based motor learning. One possible way to implement competitive physical interaction is rendering virtual springs with negative stiffness between the subjects such that they push each other away while following the same target. It is also possible to facilitate competitive physical interaction by assigning conflicting goals while subjects haptically interact with each other [48, 68].
Cooperative interaction in a teacher-student scenario can also be helpful for individual motor learning, especially for less-skilled subjects who are not able to complete the task alone. One way to obtain this interaction is by implementing a virtual uni-directional spring where force is transmitted only to the student. For a perfect teacher, this interaction resembles assistive robotic training strategies [69,70,71].
Due to their relevance to different robot-aided physical therapy strategies, implementation of competitive, collaborative, and cooperative physical interaction in a single study to compare the effects on individual motor learning could be very beneficial to our understanding of HRH interaction. We believe that such results could be further translated to robotic controllers that can mimic human-like adaptability.
Effects of partner characteristics
The studies analyzed here were focused on either effects of partner skill level on task performance/individual motor learning or effects of partner relationship on exercise intensity and motivation for rehabilitation scenarios.
Limited and conflicting results about the effects of partner characteristics on individual motor learning
Only three of the included studies investigated the effects of the training partner’s initial skill level on individual motor learning [6, 8, 37]. All of them implemented collaborative physical interaction between the subjects during planar tracking/reaching tasks under visuo-motor disturbances [8, 37] or force field [6].
Kager et al. [8] and Mireles et al. [6] reported qualitative trends and statistically significant results respectively, suggesting training with a partner that has a similar initial skill level improves individual motor learning more compared to training with an expert. The reason behind these results might be that novice partner does not need to work hard when they are connected to an expert, which can negatively impact their learning of the task. On the other hand, Beckers et al. [37] found that as the initial skill level of the training partner increases, individual motor learning increases. While the hypothesis behind this result might be that having a better teacher is better for learning new tasks, it is worth noting that it was a collaborative task with no explicit roles (e.g., teacher-student) given to the subjects. Another possible explanation is the fact that initially less-skilled partners, who are more likely to interact with better partners, have more room to improve their individual performance than initially more skilled partners. Another important difference between these studies is that Beckers et al. [37] use the same trajectory for the dyadic and solo trials with only different starting points. Therefore, cognitive memory might play a role in their results. On the other hand, Kager et al. [8] and Mireles et al. [6] use different trajectories or random targets where results are more generalizable to different motions and depend less on the properties of the selected trajectory. Moreover, while Beckers et al. [37] measured motor adaptation, Mireles et al. [6] and Kager et al. [8] investigated mostly skill learning. We believe more work on this topic is warranted with different tasks, different disturbances, and more variety on the initial skill differences of the subjects to better understand the reasons behind the conflicting results.
Transferring results to rehabilitation environments
Long-term effects need further investigation
Physical rehabilitation is usually a process that takes weeks, months, or even years. Among selected studies, only five of them implemented experimental setups where training takes place in multiple sessions during different days or weeks [5, 6, 22, 23, 45]. Moreover, only Mireles et al. [6] implemented physical interaction and had a post-retention session on a different day from the last training session. Extending the number of studies involving physical dyadic training over a long period of time and investigating long-term retention could inform the application of results to rehabilitation scenarios.
Independent investigation of partner’s nature and environmental settings is missing
In rehabilitation settings, in addition to the partner’s nature, environmental settings can influence performance and motor learning. In the studies of Gorsic et al. [3, 22], more positive results in terms of motivation and exercise intensity were obtained for participants interacting with a friend/relative at home compared to interacting with a stranger (another patient or therapist) at a clinic. However, it is worth noting that whether the increased motivation was due to interacting with a relative/friend or due to the environmental factors was not analyzed. Additional studies that independently investigate the effects of the partner’s nature and environmental settings are needed. Results from these studies can be translated to possible home or clinic-based dyadic rehabilitation.
Lack of studies on the effects of partner’s age and impairment
While there are studies analyzing different kinds of tasks, skill levels, and interaction environments, no study examined subjects’ ages among the selected references. We believe this factor should also be investigated in future studies considering that robotic haptic guidance results in different motor learning improvements among different ages [72, 73]. Another factor that might influence preference and motivation is the participants’ level of impairment. Clinical trials need to include more participants with different levels of impairment for a systematic investigation. Lessons learned regarding these factors can be applied in clinical scenarios where patients interact with each other during their training sessions.
Limited variation of interaction condition
Even though the effects of interaction type, interaction mode, and partner characteristics are investigated in different studies, only a couple of studies examine more than one of these three components with the same experimental devices and procedure. Moreover, as shown in Fig. 3, there is no single study that systematically varies each interaction component. To have more generalizable results on task performance and individual motor learning, we believe that interaction types, characteristics, and modes should be systematically changed, and the effects of a wide variety of interaction conditions should be compared.
Does physical HH interaction improve task performance?
Many studies clearly conclude that collaboratively and physically interacting with a better partner results in better task performance than interacting with a worse partner for different tracking, reaching, or trajectory following tasks at upper limb with different physical interaction characteristics (i.e., stiffness, damping). This is not surprising due to the nature of collaborative physical interaction where the partner’s individual performance directly affects the dyadic task performance.
Interestingly, some studies found that even connecting with a worse partner can improve task performance [2, 7]. Subjects using haptic forces to estimate their partner’s target to improve their own prediction of the target can be one explanation for that [11]. It is also possible that physical interaction corrects the “irregular or erratic tracking behaviours” of subjects [7]. Implicit role specialization [1] between subjects might be another explanation for the improved task performance compared to solo even if the partner’s skill level is worse.
Does physical HH interaction improve individual motor learning?
Unlike task performance, the results obtained on the effects of HH interaction on individual motor learning are more limited, contradictory, and hard to generalize. Three studies [2, 7, 37] that compared collaborative dyadic training with solo training did not report consistent results with each other. Ganesh et al. [2] and Beckers et al. [37] implemented an almost identical experimental protocol for a target tracking task with visuo-motor rotations. While Ganesh et al. [2] reported better individual motor learning with dyadic training than solo training, Beckers et al. [37] did not find any significant difference between the individual motor learning of the dyadic group or solo group.
One possible explanation for the different results might be the fact that different robotic devices were used. Neither article reported their haptic transparency or virtual environment rendering fidelity during solo and connected trials, respectively. It is important that interaction forces are rendered with high accuracy to correctly compare the results from different interaction conditions. For example, if parasitic forces felt by the user in solo mode (e.g., frictional and inertial forces) are significant compared to the interaction forces in the connected mode, the comparison of solo and connected training can be misinterpreted. Similarly, if the stiffness felt by the users is not consistent during dynamic motion, the effects of the physical interaction might be misinterpreted.
Due to the conflicting and limited number of studies, we believe further systematic studies focusing on individual motor learning are warranted. Moreover, presenting haptic transparency, and rendering performances can be beneficial to prove solo and dyadic conditions are implemented properly.
Human-like robotic controllers
Further analysis of how humans move when they are coupled to others, and transferring this knowledge to human-robot systems, can lead to advanced human-like robotic controllers. Human-like robot controllers based on improving motor learning could have significant implications on rehabilitation robotics to achieve or perhaps improve the results obtained by conventional human therapists. Moreover, knowledge of HH interaction on dyadic task performance can be used to develop robots that work together with humans in industry settings, such as carrying a large object or performing assembly tasks together.
We found several studies that did not meet our inclusion criteria, but which did investigate HH interaction to develop human-like robotic controllers or to understand dyadic communication. One way to represent human motion is to model it as an optimal controller minimizing a cost function that penalizes error and effort [11, 68, 74, 75]. Another less commonly used method is to represent the human as a first- or second-order system with delay [76, 77]. It is also important to take specialization, role sharing, and negotiation into account while modeling human behavior. Accelerator/decelerator [1], executor/conductor [78], and leader/follower [68] are some of the observed roles in human-human studies. In line with these results, Kucukyilmaz et al. implemented a haptic role exchange mechanism on a dynamic human-robot joint manipulation task and showed that the role exchange mechanism improves task performance and joint efficiency of the partners [79, 80].
Three studies used HH interaction-derived robot controllers to substitute one of the peers [11, 81, 82]. In all of these studies the subjects either did not realize that they were interacting with a robot [81, 82] or their task performance or individual motor learning results with robot partners were similar to their results with human partners [11]. However, it is worth noting that it was not clear if substituting one of the peers had an effect on the motivation because the experimental procedures did not involve any social interaction component. We believe that these results are promising for obtaining human-like robotic controllers, and they should be supported with additional studies with different interaction conditions and tasks.