Robot-Mediated Physical Human-Human Interaction

• Robot-mediated physical human-human interaction is a framework where robots serve as physical mediums to transmit forces, guidance, and social cues between humans.
• It integrates control methods like impedance shaping and virtual coupling with sensory feedback to maintain human adaptability while adding precision, repeatability, and safety.
• Applications span neurorehabilitation, social coordination, and collaborative manipulation, highlighting challenges in bidirectional control, transparency, and multi-user stability.

Robot-mediated physical human-human interaction is a framework in which one or more robotic devices are used to render specific interaction dynamics between individuals, so that the robot functions not primarily as an autonomous partner but as a physical communication medium between people. In neurorehabilitation, the canonical case is therapist-patient interaction, but the same concept extends to peers, family members, teachers and learners, and more general dyadic or networked settings. Across the literature, the central motivation is consistent: preserve therapist adaptability, tactile guidance, and human social interaction while adding robotic strength, precision, measurability, repeatability, safety, and remote reach (Vianello et al., 23 Jul 2025).

1. Definition, scope, and conceptual boundaries

Robot-mediated physical human-human interaction is distinct from both direct manual therapy and standard physical human-robot interaction. In direct manual therapy, therapist and patient exchange forces without an engineered intermediary. In standard pHRI, the human interacts with the robot as the primary partner. In robot-mediated interaction, by contrast, the robot is the enabling substrate through which one human influences another’s movement in real time. The position paper on neurorehabilitation defines the paradigm as one in which at least one individual is a patient physically connected to a robot, whose movements are influenced in real time by another user; more broadly, it frames the robot as a seamless facilitator of natural human-human interaction rather than a replacement therapist or a stand-alone coach following preprogrammed trajectories (Vianello et al., 23 Jul 2025).

This boundary matters because the design objective changes. The problem is no longer only safe compliance or accurate trajectory tracking. It becomes the preservation and regulation of a human-human channel carrying movement intent, assistance, resistance, correction, proprioceptive input, and social meaning. The same literature also distinguishes RMPHHI from bilateral teleoperation and shared control. Bilateral teleoperation typically prioritizes master-slave transparency to remote environments, whereas robot-mediated interaction prioritizes interpersonal physical communication, motor learning, and role negotiation. Shared control allocates authority between a person and a robot; robot-mediated interaction instead forms a distributed human-human motor dyad through robotic interfaces (Vianello et al., 23 Jul 2025).

A recurrent misconception is that all robot-mediated interaction must be bidirectional. The literature explicitly distinguishes bidirectional interaction, in which both users perceive interaction forces, from unidirectional interaction, in which only one user is rendered haptic feedback while the other influences the movement without receiving force feedback. The unidirectional therapist-to-patient exoskeleton framework for gait training is a concrete example of the latter, and the rehabilitation position paper treats both forms as legitimate members of the paradigm (Amato et al., 2024).

2. Interaction configurations and taxonomic structure

A unified taxonomy has emerged around how the humans are connected, how influence flows, and what task structure the coupling supports. The rehabilitation position paper organizes robot-mediated interaction by configuration, directionality, interaction medium, partner characteristics, interaction scenario, and the distinction between task performance and motor learning (Vianello et al., 23 Jul 2025).

Dimension	Main categories	Functional implication
Configuration	task-space, single-joint, multi-joint, cross-limb, guided interactions	determines what mechanical variables are exchanged
Directionality	bidirectional, unidirectional	determines whether both users perceive forces
Scenario	collaborative, cooperative, competitive	determines whether roles are symmetric or distinct

Task-space interactions exchange endpoint forces through a robot end-effector. Single-joint interactions couple one degree of freedom, often for focused rehabilitation. Multi-joint interactions, typically implemented with exoskeletons, support coordinated limb movement. Cross-limb interactions allow, for example, a therapist upper limb to influence a patient lower limb. Guided interactions allow one person to influence robot behavior indirectly, for example through kinematic demonstration. These distinctions are not merely classificatory; they determine what can be communicated physically. Endpoint coupling is effective for directional guidance, whereas explicit postural correction often requires multi-joint or exoskeletal coupling (Vianello et al., 23 Jul 2025).

The interaction medium itself is commonly described as a virtual spring-damper. In task space this can be written as

$F = K(x_1 - x_2) + D(\dot{x}_1 - \dot{x}_2),$

and in joint space as

$\tau = K(q_1 - q_2) + D(\dot{q}_1 - \dot{q}_2).$

The same framework also supports asymmetric coupling with therapist and patient stiffness and damping parameters $K_T, K_P, D_T, D_P$. This asymmetry is clinically important because it allows strong haptic information to the therapist, reduced therapist effort, and individualized workload distribution (Vianello et al., 23 Jul 2025).

Partner characteristics further modulate the interaction. The literature explicitly notes that relative skill, confidence, familiarity, relationship type, and the number of users influence effort, engagement, and emergent roles. This suggests that the medium cannot be treated as purely mechanical. Social and relational structure is part of the control problem, not merely part of user evaluation (Vianello et al., 23 Jul 2025).

3. Control architectures and mediation mechanisms

The dominant control substrate in this area is impedance or admittance shaping, often augmented by mechanisms that let one person alter not just current motion but future robot behavior. A particularly clear example is physically interactive trajectory deformations, added on top of impedance control. For a future desired segment $\gamma_d$, the deformed segment is

$\tilde{\gamma}_d = \gamma_d + \mu \delta H f_h,$

so that human-applied force modifies a bounded future motion segment rather than only the instantaneous tracking error. The method was introduced for single-human pHRI, not for two humans, but the paper explicitly notes that it can be viewed as a building block for transmitting one human’s intent through a robot to another human (Losey et al., 2017).

A second line of work implements virtual coupling directly between the humans. In unidirectional gait training, therapist IMU-derived joint angles become desired exoskeleton joint angles, and the patient-side torque is rendered by

$\tau_{\text{des}}(s) = Z(s)\,\theta_e(s),$

with

$$Z(s) = k_{\text{high\_f}} \frac{s + f_{\text{cut}}\dfrac{k_{\text{low\_f}}}{k_{\text{high\_f}}}}{s + f_{\text{cut}}}.$$

This interaction law is stiff at low frequencies and less stiff at high frequencies, so slow therapist-patient mismatch is corrected strongly while rapid fluctuations are attenuated. The paper treats this frequency shaping as central to balancing guidance authority and compliant safety (Amato et al., 2024).

Object-mediated interaction uses the same logic at payload level. In collaborative manipulation of a cable-suspended load by multiple aerial robots and a human, the payload reference is shaped by a 6D admittance law,

$$\mathbf{M}\ddot{e}_{adm} + \mathbf{D}\dot{e}_{adm} + \mathbf{K} e_{adm} = \begin{bmatrix} F_H \ M_H \end{bmatrix},$$

so human-applied force and moment produce translational and rotational payload motion. Because the lower-level robots only need to realize the resulting payload wrench, the human interacts with an apparent virtual object rather than with raw multi-robot dynamics (Li et al., 2022).

Force consistency is another recurring issue. For tethered aerial guidance, a robot-state-only controller loses effective guiding force as the human walks faster; adding human velocity feedback yields

$u_A = K e + \bar f + B_A \dot p_H,$

and the paper proves that the steady-state cable force then satisfies $f_c = \bar u$, independent of human walking speed, under the stated damping condition. The relevance to mediation is explicit: if a mediator ignores the receiver’s motion state, the same intended cue will feel different to different responders (Allenspach et al., 2022).

More generally, partner-aware control formalizes the interacting bodies as a coupled dynamical system rather than treating external forces as disturbances. In the humanoid formulation, interaction wrenches are mapped into task-space contributions, and helpful interaction is exploited rather than canceled. This is a foundational idea for mediated human-human interaction because one human’s assistance should not be washed out by the robot if it is task-relevant and safe (Tirupachuri et al., 2018).

A newer communication-oriented architecture adds a non-haptic layer. BRIDGE represents a physical assistive trajectory as

$\tau = K(q_1 - q_2) + D(\dot{q}_1 - \dot{q}_2).$0

and lets users modify position, velocity, and force in real time through natural language, with robot verbal assurances or clarifying questions. The paper is single-user and assistive rather than explicitly interpersonal, but it makes transparency part of the interaction controller rather than a separate user-interface layer (Wang et al., 15 Jan 2026).

4. Sensing, estimation, and the interpretation of human state

Robot-mediated interaction depends not only on rendering forces but also on inferring body state, intent, and comfort from incomplete measurements. Several strands of work address this sensing problem at different scales.

At local body-contact scale, multidimensional capacitive servoing uses a six-electrode capacitive array and a temporal neural estimator to infer a 4D local limb pose and servo the robot along the body. The feedback controller is

$\tau = K(q_1 - q_2) + D(\dot{q}_1 - \dot{q}_2).$1

with additional constant advance along the local $\tau = K(q_1 - q_2) + D(\dot{q}_1 - \dot{q}_2).$2-axis, allowing the robot to move proximally or distally along a limb while regulating offset and orientation. This is not robot-mediated human-human interaction by itself, but it provides the near-body perception layer needed for mediated touch, bathing, dressing, and limb guidance (Erickson et al., 2021).

At whole-posture scale, ergonomically intelligent pHRI shows that the interaction-point trajectory sensed by the robot can be sufficient to estimate human upper-body posture. The observation model is

$\tau = K(q_1 - q_2) + D(\dot{q}_1 - \dot{q}_2).$3

and the resulting robot-only estimate achieved median deviation less than $\tau = K(q_1 - q_2) + D(\dot{q}_1 - \dot{q}_2).$4 rad from motion capture. The same work introduces DULA, a differentiable ergonomic surrogate with $\tau = K(q_1 - q_2) + D(\dot{q}_1 - \dot{q}_2).$5 accuracy comparing to RULA, then uses it in postural optimization. This suggests that a mediator can simultaneously act as motion interface, posture sensor, and ergonomic optimizer (Yazdani et al., 2021).

At whole-body dynamics scale, wearable-sensor fusion has been used to estimate link accelerations, external/contact forces, internal joint forces, and joint torques in real time. The thesis formulates human dynamics estimation as a MAP problem,

$\tau = K(q_1 - q_2) + D(\dot{q}_1 - \dot{q}_2).$6

rather than as deterministic inverse dynamics. A plausible implication is that mediation need not rely only on endpoint forces; it could exploit estimates of effort distribution and internal loading when deciding how to transform one human’s action into another’s experience (Latella, 2019).

Intent inference has also been studied through force-kinematic features. In collaborative object manipulation, the key feature is interaction power,

$\tau = K(q_1 - q_2) + D(\dot{q}_1 - \dot{q}_2).$7

and, for symbolic goal inference, projected power toward candidate goal directions,

$\tau = K(q_1 - q_2) + D(\dot{q}_1 - \dot{q}_2).$8

The study argues that humans negotiate in epochs rather than through continuously smooth signals, and that temporary conflict carries information. This suggests that a robot mediator should preserve or interpret force-based negotiation rather than simply averaging it away (Rysbek et al., 2023).

Finally, social perception remains relevant even when the robot is “just” a medium. In embodied pHRI, Warmth and Competence were the strongest predictors of which robot behavior people preferred to experience again, outperforming the tested Godspeed dimensions. A plausible implication is that a mediation robot must be perceived as both socially considerate and physically capable if it is to be accepted in therapist-patient, caregiver-patient, or socially sensitive contexts (Scheunemann et al., 2020).

5. Application domains and empirical evidence

Neurorehabilitation is the most explicitly developed application domain. The rehabilitation position paper argues that RHHI can reintegrate therapist adaptability and tactile expertise into robotic therapy while adding offloading, measurement, and repeatability (Vianello et al., 23 Jul 2025). Concrete systems already instantiate parts of this agenda. In therapist-robot-patient remote haptic control, thirty-two participants compared haptic demonstration with visual demonstration in an arm-pose guidance task, and the abstract reports that haptic demonstration significantly reduced movement completion time and improved smoothness while yielding fewer verbal instructions and greater trainer competence (Luciani et al., 25 Feb 2026). In the unidirectional gait-training framework, healthy teacher-student experiments showed emergence of synchronization, transmission of asymmetric gait patterns, and “virtual obstacles” conveyed through haptic coupling without predefined patient trajectories (Amato et al., 2024).

Outside rehabilitation, close-contact social coordination has been studied through a humanoid leading a slow waltz. The whole-body controller combines task-space admittance and impedance, and the main empirical result is that gentle haptic cues plus simple audio timing support are preferred to stronger displacement cues. The best overall trial was Hand Wrench plus Step Count, whereas Hand Displacement plus Step Count was least preferred. This is directly relevant to mediated interaction because it shows that guidance quality is inseparable from comfort, and that stronger cues can improve clarity while degrading acceptance (Charbonneau et al., 2024).

Object-centered mediation has also been demonstrated with aerial robots. Multi-robot collaborative transportation and manipulation supports human interaction with a rigid-body payload in all 6 DoF using force-sensor-free external wrench estimation, 6D admittance, and redundancy-based separation constraints. The real system allowed approximately $\tau = K(q_1 - q_2) + D(\dot{q}_1 - \dot{q}_2).$9 m translation in $K_T, K_P, D_T, D_P$0 and $K_T, K_P, D_T, D_P$1, $K_T, K_P, D_T, D_P$2 m in $K_T, K_P, D_T, D_P$3, $K_T, K_P, D_T, D_P$4 roll and pitch, and $K_T, K_P, D_T, D_P$5 yaw under human guidance, while the optimization-based safety controller maintained the reported human-robot and inter-robot separations (Li et al., 2022).

At the dyadic whole-body level, the dual-humanoid framework Harmanoid shows what interaction-aware motion transfer looks like when contact geometry is modeled explicitly. Its contact-aware retargeting and interaction-driven controller raised success from $K_T, K_P, D_T, D_P$6 for ExBody and $K_T, K_P, D_T, D_P$7 for HOVER to $K_T, K_P, D_T, D_P$8 on the reported dual-interaction benchmark. This is not yet robot-mediated human-human interaction in deployment, but it provides a controlled proxy for physically grounded, socially meaningful whole-body interaction that single-agent imitation pipelines fail to capture (Liu et al., 11 Oct 2025).

Across these domains, a consistent empirical pattern appears: mediation benefits from compliance, explicit modeling of relational structure, and some form of transparency about what the robot is doing. The BRIDGE study illustrates the last point. Older adults successfully modified assistive trajectories in real time through speech in scratching, feeding, and bathing tasks, and bidirectional verbal feedback significantly increased perceived interactivity, grounding, and transparency over a unidirectional ablation, even though both modifiable systems outperformed baseline on perceived control and task success (Wang et al., 15 Jan 2026).

6. Open problems, controversies, and research directions

The field remains technically fragmented. Several building blocks exist, but many papers explicitly note that they are not complete mediation frameworks. Physically interactive trajectory deformation allows one human to shape future desired motion, but it does not address two simultaneous human inputs, conflict arbitration, or multi-user stability. The paper itself states that what is missing for direct transfer is a multi-user arbitration and stability framework (Losey et al., 2017).

Directionality remains a central design controversy. Unidirectional systems are cheaper and easier to deploy, and they already transmit timing, coordination, asymmetry, and obstacle-related cues. Bidirectional systems, however, are repeatedly associated with embodiment, richer therapist perception, and more natural physical communication. The rehabilitation position paper does not resolve this tradeoff; instead it frames asymmetric interaction as a promising compromise (Vianello et al., 23 Jul 2025).

Transparency and safety also pull in opposite directions. Admittance, frequency shaping, and null-space safety motion help, but several papers note unresolved issues around passivity, communication delay, actuator saturation, cable slack, and multi-human closed-loop stability. The gait-training framework provides torque slew limits and ROM saturation but no explicit delay analysis; the aerial payload system uses quasi-static wrench estimation and no formal passivity proof; the language-mediated BRIDGE system pauses motion for interpretation, which may be acceptable in bathing-like tasks but not in dynamic co-manipulation (Amato et al., 2024).

Clinical validation is still limited relative to the breadth of the paradigm. The neurorehabilitation position paper explicitly calls for larger trials, direct comparisons against conventional manual therapy and existing robot-assisted therapy, standardized but customizable exercise libraries, therapist training protocols, and broader functional tasks beyond stereotyped reaching. It also highlights cost, accessibility, workflow integration, remote networking, privacy, and therapist adoption as unresolved translational barriers (Vianello et al., 23 Jul 2025).

A final misconception is that better physical transparency alone will solve the problem. The broader literature argues otherwise. Human-human physical interaction is simultaneously mechanical, perceptual, and social. Warmth and Competence predict preference in embodied pHRI; clarifying questions materially improve interaction transparency in assistive language-mediated control; and dance-like close-contact interaction shows that gentle, socially legible cues can outperform stronger but less acceptable ones. This suggests that future robot-mediated physical human-human interaction will need to unify mechanics, state estimation, arbitration, and social communication rather than optimizing any one of them in isolation (Scheunemann et al., 2020).