Active Perception of Upper-Limb Functional Mobility through Human-Robot Handovers

Abstract

The reintegration of people in the workplace after an upper-limb injury is often a long process. Collaborative robots and humanoids operating in shared workspaces must adapt their behaviour to the physical capabilities of their human partners. Achieving this requires robots to autonomously estimate task-relevant human mobility during interaction, without relying on external clinical assessments. In this work, we formulate the estimation of upper-limb functional mobility as an active perception problem. We model task-available upper-limb functional ranges of motion through impairment coefficients that parameterise reductions of healthy joint limits, and integrate this representation into a human model used by the robot during interaction. Based on this formulation, we propose an online active perception pipeline in which the robot actively probes the user through repeated handover requests while progressively updating a Bayesian belief over the impairment coefficients. This enables the robot to continuously refine its internal model of the user’s joint-level capabilities through interaction rather than passive observation. The proposed framework was evaluated through simulation studies and a proof-of-concept physical validation with a robotic platform and enacted human impairments, demonstrating that the pipeline can accurately recover the impairment coefficients during real human-robot handovers.

I INTRODUCTION

Going back to work after having recovered from an upper-limb injury is a delicate moment for workers, because they might have not completely retrieved their full mobility prior to the accident. Collaborative robots and humanoids are progressively appearing in industrial environments and could facilitate this process by adapting their behaviour to the physical capabilities of individual workers [1, 2]. Such adaptations require robots to estimate the user’s residual functional mobility. With such knowledge, robots can further tune the physical interaction with the user to make their reintegration in the workplace more manageable and personalised.

Additionally, robots have the potential to plan their own motions and collaborate with workers to avoid compensatory movements and take over the workers’ more proper motions related to work activity, because their functional mobility cannot yet be fully reengaged [3, 4]. Although robotic solutions that can account for compensatory movements have been developed in rehabilitation settings [5], a system capable of achieving the same purpose in industrial settings has yet to be designed. Indeed, the few existing works addressing upper-limb impairments in human-robot collaboration primarily substitute the user’s movements with robotic assistance or minimise the need for arm motion altogether [6, 7, 8], rather than explicitly modelling and adapting to the user-specific joint level limitations.

In order to adapt an interaction to a worker’s physical capabilities, the robot requires a quantitative representation of the user’s mobility at the joint level. A joint’s range of motion (ROM) describes its overall admissible angular interval. However, the portion of that interval available or expressed during a goal-directed activity may be smaller. We refer to this task-available interval as the functional range of motion (fROM). In clinical settings, scores like the Fugl-Meyer’s [9] are used to give a qualitative evaluation of the user’s ROMs. However, when it comes to human-aware motion planning, clinical scores lack in providing quantitative linking to the mobility at the joint level [10], which could be used by a robot to properly tune the interaction according to the specific joint movements to elicit [11]. Furthermore, clinical scores are measured while having patients performing unnatural movements, which could alter the functional ROM (fROM) during more usual work activity. In relation to the upper limb, previous attempts of measuring its ROMs have already been performed in literature. Keyvanian et al. [12] explored an unsupervised learning approach to learn ROM boundary functions, by collecting full-range movements of healthy people, exploring one by one all the possible ROMs of the upper arm. They then defined an impairment index, a high-level quantitative metrics of the user’s impairment that takes into account all the ROMs estimated altogether. However, in this approach, they explored free movements of the arm without focusing on goal-oriented and manipulation tasks [13], having participants extensively move each joint at a time to their fullest extent. Considering the upper limb, the engagement of each degree of freedom (DOF) of it changes from using them one at a time and altogether within a goal-oriented action [14]. Therefore, the fROMs of the user acting in a setting may differ. Because of this, Liu et al. [15] proposed a framework that leveraged information from task-based recordings of people with different ROMs. Upon receiving clinical assessment scores of the ROMs related to the recordings [9], they learned a mapping from the clinical scores to the ROMs, so that the robot could learn the fROMs of a new user by using their clinical scores and without motion tracking. Although effective, this methodology relies on having the clinical scores available and updated. While these could be available at the beginning of the reintegration in the workplace, it can be demanding to summon a clinician again in the future to track and monitor the progress of the recovery outside of the clinical settings. A robot should be able to evaluate the ROMs of the user independently, without having to rely on external human inputs for this information.

Therefore, in this work, we developed a pipeline through which a collaborative robot estimates the impairment coefficients associated with the user’s upper-limb fROM through an active perception process (see Fig. 1). In this, the robotic system repetitively hands over an object to the human user. Based on the user’s kinematics at each handover pose, the belief over the impairment coefficients is updated. The resulting coefficients parameterise the modeled fROM bounds of the considered joints. This process is repeated until the system finds an accurate estimate of the fROMs inspected. No previous learning is requested to perform this estimation. We demonstrated the validity of this approach through an experimental validation in which a participant performed the requested handovers while having their upper arm movements artificially hindered. The contributions of this work are threefold:

• a robot-centric formulation of functional upper-limb mobility through impairment coefficients;
• an online active perception pipeline where the robot initiates handovers and updates a belief over the user’s functional mobility;
• validation of the proposed framework through deployment on a real robotic platform interacting with a participant with enacted impairments.

This paper is structured as follows. Upon having introduced our work (see Section I), Section II describes the human modeling adopted for the problem and the active perception pipeline used for the ROM perception problem. Section III and IV report the experiments carried out and the related results, respectively. Finally, Section V and VI reason about the interpretations of the results and summarise the content of the paper.

II MATERIALS AND METHODS

The ROM active perception process was structured as follows. The robot moves the object to hand over to a designated pose in space and holds it in place, waiting for the user to reach out for it. While doing so, the robot makes an internal estimate of the user’s kinematics at the handover point based on the current handover pose of the object and estimate of the user’s fROMs of interest. When their hand reaches the closest possible to the handover point, the kinematics of the user gets recorded. This recording and the robot’s estimate of the human kinematics are used to update the current belief on the fROMs of interest. This process is repeated until a condition on the estimate is met (see Subsection II-C) and the perception process terminates. The pipeline is reported in Fig. 2 and its components are described below.

II-A Problem definition

In our design, the user’s ROMs are modeled as in [11]:

\displaystyle\bm{q} \\ \displaystyle\in\left[\bm{q}_{\min}+\Delta\bm{q}_{\min},\;\bm{q}_{\max}-\Delta\bm{q}_{\max}\right],

where \bm{q}\in\mathbb{R}^{N_{j}} are the joint values of the person’s kinematic chain of the arm, with N_{j} DOFs considered, \bm{q}_{min} and \bm{q}_{max} respectively the lowest and greatest values for healthy joints and

\displaystyle\Delta\bm{q}_{\min} \\ \displaystyle=\bm{\alpha}\left(\min\left(\bm{q}_{n}-\zeta,\bm{q}_{t}\right)-\bm{q}_{\min}\right), \\ \displaystyle\Delta\bm{q}_{\max} \\ \displaystyle=\bm{\alpha}\left(\bm{q}_{\max}-\max\left(\bm{q}_{n}+\zeta,\bm{q}_{t}\right)\right),

where \bm{q}_{n} are the joint values of the arm in their nominal configuration (\zeta>0 allows small changes in case of full impairment), \bm{q}_{t} are the measured joint values and \bm{\alpha}\in[0,1]^{N_{j}} is the vector of impairment coefficients (ICs). Here, the range in (1) represents the healthy ROMs of the model, whereas the bounds resulting from (2) represent its modeled fROMs under the impairment configuration \bm{\alpha}. The ICs represent the level of impairment of each joint of the considered kinematic chain of the arm, with 0 representing full mobility and 1 full impairment. In other words, \bm{\alpha} reduces the healthy ROM bounds to obtain a model-based representation of the user’s task-available fROM. Our active perception process aims to estimate the components of \bm{\alpha} for the DOFs of interest, therefore an estimate vector \hat{\bm{\alpha}} with size N_{imp}\leq N_{j}. In this work, we focused on shoulder and elbow flexion because coordinated shoulder flexion and elbow motion are the principal contributors to forward reaching [16], which is the movement elicited by the proposed handover task. Accordingly, the proposed pipeline does not directly observe the boundaries of each joint ROM. Instead, it estimates the ICs that parameterise the functional joint limits represented by the human model.

II-B Human model

The human model considered in this work is reported in Fig. 3 [11]. The model incorporates one IC for each considered DOF. These coefficients reduce the corresponding healthy ROM bounds according to (2), thereby defining the modelled fROMs. Indeed, the model is capable of computing its own inverse kinematics (IK) by taking into account the ICs. Besides, to estimate the configuration of the upper limb, given an impairment and a goal pose, we run an underlying optimisation process from [4], within which the ICs are incorporated, as well. In Fig. 2, this computation is represented by the IK Human Predictor block. In particular, the optimisation solves

\displaystyle\min_{\boldsymbol{q}} \\ \displaystyle\left\|\bm{A}(\boldsymbol{q}-\boldsymbol{q}_{t})\right\|^{2}+\gamma\left\|(\mathbf{I}-\bm{A})(\boldsymbol{q}-\boldsymbol{q}_{n})\right\|^{2}, \\ \displaystyle\text{s.t. }(\ref{eq1}),\boldsymbol{x}_{obj}(\boldsymbol{q})=\boldsymbol{h},

where the first term minimises the deviation of fully or partially impaired joints from their current configuration. This is achieved through the diagonal weighting matrix \bm{A}=\mathrm{diag}(\bm{\alpha}), whose entries correspond to the ICs and assign larger penalties to the motion of more severely impaired joints. The second term instead penalises compensatory movements by minimising the deviation of the healthy joints from the nominal comfortable configuration \bm{q}_{n}. The scalar parameter \gamma\in\mathbb{R} weights this contribution, giving priority to preserving the current configuration of impaired joints while softly encouraging comfortable postures for the unconstrained joints.

The task-specific constraints enforce that the object pose \boldsymbol{x}_{obj}, which corresponds to the grasp pose reached by the user’s hand during the interaction, matches the desired task pose. Within the proposed active perception framework, the desired task pose \boldsymbol{h} is iteratively updated to explore different handover configurations, thereby eliciting informative upper-limb motions that improve the estimation of the user’s impairment coefficients.

This model was chosen under the strong assumption that only the movements of the upper arm were relevant during an activity to determine the ROMs of its joints. In real-life settings, the active perception process needs to be determined outside the regular working activity, to enable reproducible conditions for the active perception process and avoid the interference of compensatory movements [4, 17]. As a consequence, it is assumed that during the active perception process the user is told to hold its position in space and to use only the movements of their arm for the reaching movements [3]. This prevents the measurements from being affected by displacement of the person and trunk compensation [11, 17].

II-C Active perception

The process iteratively estimates \hat{\bm{\alpha}} via Bayesian updates through a particle filter [18], computed in the Belief Updater block (see Fig. 2). This choice for the core estimation algorithm was made because \bm{\alpha} does not vary over time during the interaction, is characterised by non-linear relationships with the user’s joint values, due to the dependence on a kinematic chain, and its components’ value lie within a well-determined range (see Subsection II-A). The particle filter represents the posterior belief over the ICs as the weighted empirical distribution

p(\hat{\bm{\alpha}}\mid\mathcal{D}_{t})\approx\sum_{i=1}^{N_{p}}w_{i,t}\,\delta(\hat{\bm{\alpha}}-\bm{\alpha}_{i,t}),

where \mathcal{D}_{t} denotes the observations collected up to iteration t. For reporting and termination purposes, this belief is summarised through its weighted mean \bm{\mu}_{\bm{\alpha},t} and marginal variance vector \bm{\sigma}_{\bm{\alpha},t}^{2} .

The robot starts the active perception requesting a handover to the user by moving the object to a handover pose \bm{h}_{t}. The system then detects \bm{q}_{t} when it finishes reaching out to the object and uses it to define a log-likelihood distribution of the IK of the user at \bm{h}_{t} as

\displaystyle\mathcal{LL}(\bm{q}_{t})=ln(\mathcal{N}(\bm{q}_{N_{o},t},\sigma_{q}^{2}\bm{I}_{N_{o}})),N_{o}\leq N_{j},

with \bm{q}_{N_{o},t} a vector of the components of \bm{q}_{t} used to define the log-likelihood upon and \bm{I}_{N_{o}} as N_{o}\times N_{o} identity matrix. In our implementation, we set \sigma_{q}^{2} to 0.2 and considered the same DOFs related to \hat{\bm{\alpha}} (see Subsection II-A). At the same time, a set of N_{p}=50 different particles \mathcal{A}_{t}=\{\bm{\alpha}_{i,t}\}_{i=1}^{N_{p}}, that is random instances of the IC vector to estimate, are defined and assigned each an equal weight of 1/N_{p}, collected in a vector \bm{W}_{t}. In the Bayesian framework, the particle set \mathcal{A}_{t} together with the associated weight vector \bm{W}_{t} represents the prior distribution of \bm{\alpha} at time t. The pipeline then computes the IK solutions of the human model for pose \bm{h}_{t} given \mathcal{A}_{t} (see Subsection II-B), defined as

\displaystyle\mathcal{Q}_{t}=\left\{IK(\bm{h}_{t}\mid\bm{\alpha}_{i,t})\right\}_{i=1}^{N_{p}}.

This produces a set of candidates IK solutions for our internal human model, given the impairment belief based on the current particles and their weights. For these joint vectors, the log-probability of each of them is evaluated and used to compute the quantity

\displaystyle\ln(\bm{W}_{t+1}) \\ \displaystyle=\bm{L}_{t}+\ln(\bm{W}_{t}+\epsilon), \\ \displaystyle\bm{L}_{t} \\ \displaystyle=\left[\mathcal{LL}(\bm{q}_{i})\right]_{\bm{q}_{i}\in\mathcal{Q}_{t}}.

\epsilon is a very small positive quantity used to avoid computing the logarithm of 0. Summing the log-likelihood terms with the logarithm of the previous weights allows them to be weighted on the weights the particles had prior to the update. Upon normalisation of ln(\bm{W}_{t+1}) in the log space and exponentiation, the vector \bm{W}_{t+1} is obtained and becomes the new weight vector associated with the particles. At this point, the belief on \bm{\hat{\bm{\alpha}}} are updated as follows:

\displaystyle\bm{\mu}_{\bm{\alpha},t} \\ \displaystyle=\frac{\sum_{i=1}^{N_{p}}w_{i,t+1}\bm{\alpha}_{i,t}}{\sum_{i=1}^{N_{p}}w_{i,t+1}}, \\ \displaystyle\bm{\sigma}_{\bm{\alpha},t}^{2} \\ \displaystyle=\frac{\sum_{i=1}^{N_{p}}w_{i,t+1}(\bm{\alpha}_{i,t}-\bm{\mu}_{\bm{\alpha},t})^{2}}{\sum_{i=1}^{N_{p}}w_{i,t+1}},

where w_{i,t+1} are the components of \bm{W}_{t+1}. After having performed the update, we check whether \bm{\sigma}_{\bm{\alpha},t}^{2}<\sigma_{thr}^{2} to to guarantee the precision of the estimate. In our design, we set \sigma_{thr}^{2}=0.003, so that the error would not be bigger than 0.1 for the majority of the times. If the condition is met for three consecutive iterations of the process just described, the estimate of \hat{\bm{\alpha}}=\bm{\mu}_{\bm{\alpha},t} is accepted. Otherwise, we generate a new handover request to the user. If the termination condition cannot be met after 25 handovers, the estimate of \hat{\bm{\alpha}}=\bm{\mu}_{\bm{\alpha},t} at that time is accepted anyway.

During the estimation, some of the particles could be assigned weights close to 0, thus not being so significant for the estimation. When many particles do not influence the estimate anymore, the elements of \mathcal{A}_{t} all undergo a resampling process. For this purpose, we define the quantity

\displaystyle N_{\mathrm{eff}_{t}} \\ \displaystyle=\frac{1}{\sum_{i=1}^{N_{p}}w_{i,t}^{2}}

as indicator of the amount of particles that have non-negligible weight, thus effective to the estimation. When the quantity N_{\mathrm{eff},t}/N_{p} is under a certain threshold (0.4 in our implementation), all the particles are resampled with replacement weighted on \bm{W}_{t}, perturbed with Gaussian noise \mathcal{N}\!\left(\bm{0},\sigma_{j}^{2}\bm{I}_{N_{imp}}\right) and assigned 1/N_{p} weights each again. In our implementation, \sigma_{j}^{2}=0.05.

III EXPERIMENTS

The active perception pipeline was evaluated in simulated and physical experiments by estimating the ICs associated with shoulder and elbow flexion [3] (referring to Fig. 3, i=2 and i=4; from this point on, we refer to them as \alpha_{S_{y}} and \alpha_{E_{y}}, respectively). The ICs of all remaining upper-limb DOFs were fixed to 0, corresponding to unrestricted healthy ROMs in the model [3]. In simulation, all possible combinations of their related ICs, ranging from 0 to 1 with a step of 0.025641, were estimated, for a total of 1600 2D vectors. In this case, the recordings in real-life were replaced by another instance of the IK Human Predictor block (see Subsection II-B) running uniquely with the configuration of ICs to estimate at that time for \alpha_{S_{y}} and \alpha_{E_{y}} and with the ICs of the remaining DOFs fixed to 0.

The real-life experiments were performed with a participant (male, age: 32, height: 1.80 m) who, additionally to the case of moving the arm freely while standing still, enacted half and full impairment of either \alpha_{S_{y}} or \alpha_{E_{y}} (see Fig. 4). To reliably reproduce the impairment on \alpha_{E_{y}}, they wore a harness that could bound it within different tunable ranges [15, 19] (see Figg. 1 and 4). In the experimental setting, a Franka Robotics® (Munich, Germany) Franka Research 3 was used for performing the handovers during the active perception process. At the startup, the space \bm{h}_{t} could be selected from a cube whose position in the workspace was tailored on the height and reachability of the user (see Fig. 5) [20, 21].

The handover poses were chosen randomly within the set space. The robot initially grabbed a mug and then moved it so that a reference frame on the handle coincided with the selected handover pose (see Fig. 5). The joint values and poses of the participants were detected through Real-Move (Real-MoveTM, Genoa, Italy), markerless motion capture system [22]. The poses were used at the startup to tailor the action space, while the joint values were processed only when the participant reached the handover pose to the closest possible extent permitted by the enacted impairment. More precisely, the skeleton joint values resulting from the circular average of the 7 recordings following the recording trigger given by the experimenter corresponded to the \bm{q}_{t} vector processed by the pipeline. The skeleton detection performed by Real-Move was compatible with the rendering of our human model (see Subsection II-B), which was scaled to the participant’s bodily features prior to beginning. While the estimate was being updated, the participant reconvened to their starting resting stance, waiting for the next handover request.

The whole ensemble was controlled through a ROS 2 Humble workspace running on Ubuntu 22.04 (12th Gen Intel® Core™ i7-12700H × 20). The IK Human Predictor block (see Subsection II-B) ran on ROS Noetic and communicated with the remaining system through a ROS bridge. The experiments were carried out at the HRI2 Laboratory, in accordance with the Helsinki Declaration, and the protocol was approved by the Comitato etico territoriale Regione Liguria (Protocol IIT_HRII_002).

IV RESULTS

The experiments in simulation were used to report statistics on the performance of the active perception pipeline, as they rely on the possibility of reiterating the experiment numerous times and under many impairment settings. Regarding the real-life recordings, instead, they relate to fewer instances, so they were used to show the progression of the active perception process over time. The results are shown below.

IV-A Simulation

Table I reports the performance of the IC estimation process (see Section III). They were reported by considering cases of low and high impairment values for the shoulder and elbow flexion set as ground truth, with 0.5 representing the threshold between low and high impairment. The error was computed as the mean of the components of |\hat{\bm{\alpha}}-\bm{\alpha^{*}}|, with \bm{\alpha^{*}} ground truth vector of the ICs to estimate. The performance of the active perception are lower in the cases of low \alpha_{E_{y}}^{*}, with the mean error being higher. Besides, the number of requested iterations is higher, relating with the termination condition met fewer times. Noticeably, the number of resamplings is lower than the cases with high \alpha_{E_{y}}^{*}.

IV-B Physical validation

Fig. 6 shows the evolution of the estimated ICs over the course of the interactions, for the impairment configurations explored (see Section III). For three of the recordings, the pipeline was able to converge to a precise estimate before reaching the highest number of iterations permitted. On the whole, the pipeline was able to understand from the first iterations the magnitude relationship between the two impairments being estimated. Fig. 7 qualitatively illustrates two iterations of the proposed pipeline, highlighting the agreement between the measured postures and the reconstructed ones using the current ICs estimate, which drives the subsequent estimate update. For \alpha_{E_{y}}^{*}=1.0, the pipeline correctly identified the presence of an elbow flexion limitation as early as the second iteration. However, the reconstructed posture still exhibited a noticeable discrepancy from the measured one. By iteration 8, the estimated posture closely matched the actual human configuration, resulting in early convergence at iteration 11 (see Fig. 6). In contrast, for \alpha_{S_{y}}^{*}=1.0, the algorithm initially underestimated the severity of the limitation. This behaviour can be explained by the fact that the handover pose selected during the early iterations primarily required elbow flexion, providing limited information about the subject’s shoulder mobility. As a result, additional observations were needed before accurately estimating \alpha_{S_{y}}, leading to convergence after 22 iterations (see Fig. 6). On the other hand, the estimate appears less accurate in the cases of low \alpha_{S_{y}}^{*} and \alpha_{E_{y}}^{*}=0. Specifically, \hat{\alpha}_{E_{y}} is off by 0.4 with respect to \alpha_{E_{y}}^{*}, while \hat{\alpha}_{S_{y}} is much more accurate at the process termination, even with different values of \alpha_{S_{y}}^{*}.

V DISCUSSION

The results demonstrate both the potential of the proposed active perception pipeline and the main limitations that currently affect its estimation performance. They are discussed below, followed by broader considerations.

V-A Estimation performance and limitations

This work presented a pipeline for actively estimating the impairment coefficients that parameterise a user’s task-available upper-limb fROM through repeated human-robot handovers. With a simulated human model with the same underlying IK optimisation (see Subsection II-B), this active perception process was capable of estimating the ICs associated with shoulder and elbow flexion with an error on average up to a tenth of the range of possible values. This was achieved on average in less than 25 handovers requests, thanks to the early termination condition (see Subsection II-C). These results were achieved without any learning previously performed, by only relying on the online estimate of the ICs through the kinematics of the user at the handover points. All these results make the pipeline easily deployable in a robot to update its internal representation of the user’s functional joint limits when needed, for a more personalised collaboration (see Section I).

Despite these promising features, several opportunities remain for improving the proposed approach. Looking at the online validations (see Subsection IV-B), the pipeline was less accurate when \alpha_{E_{y}}^{*} was set to 0 and \alpha_{S_{y}}^{*} was low. This can be explained by the following consideration. To understand that the elbow is not impaired, the pipeline would need to elicit configurations spanning a sufficiently large portion of the elbow’s available fROM to perform the right estimate under this setting. As the shoulder is a more proximal joint in the arm kinematic chain, its only movement allows a person to greatly reduce the distance from their hand to the target point. When this happens, less movement is required by the elbow to reach the target. This means that the pipeline did not manage to elicit flexion of the elbow through the handovers requested and did not elicit enough elbow excursion to constrain the corresponding IC accurately when the shoulder was free to move. This observation is confirmed by the results achieved in simulation (see Subsection IV-A), that show similar performance decrease with low \alpha_{E_{y}}^{*}.

V-B Future developments and deployment on humanoids

The limitations introduced by the current action space could be addressed by modifying its geometry so as to elicit upper-limb configurations requiring a larger contribution from elbow flexion [16]. The action space adopted in this study (see Fig. 5) may therefore be too limited, and wider or more complex spatial regions may be needed. Any such redesign would also have to account for the reachable workspace of the robotic platform, since not all informative target poses may be physically accessible.

Secondly, the handover poses used during the estimations were randomly selected from a designated space (see Section III). Future work should investigate an information-driven policy that selects the next handover pose according to the current belief over the user’s ICs. Such a policy could maximize the expected information gained from each interaction and reduce the number of handovers required to obtain an accurate ROM estimate. Secondly, the experimental validation focused on shoulder and elbow flexion, although the human model includes ICs for additional upper-limb DOFs (see Subsection II-B). Extending the estimation framework to infer a broader set of ICs would provide the robot with a richer representation of the user’s functional capabilities, enabling more personalized human-robot collaboration [11]. Finally, the current upper-limb model considers a kinematic chain with 7 DOFs (see Subsection II-B). Extending the model to include additional DOFs could increase the information extracted from each handover and improve the fidelity of the estimated mobility profile. However, this extension would also require reliable sensing of the additional joint variables included in the model.

Beyond the specific handover implementation, a broader conceptual contribution of this work lies in framing fROM estimation as an active perception problem based on guided reaching. The proposed formulation is not tied to a specific robot platform: it requires an embodied system capable of presenting target poses that elicit informative upper-limb configurations and of observing the resulting human motion. From the interaction perspective, the robot acts as a spatially situated target presenter, while its own kinematic constraints determine which handover poses can be physically realised. Consequently, if a desired pose is unreachable, the active perception process can remain applicable by selecting an alternative reachable pose with comparable information value.

This formulation is particularly relevant for humanoid robots and mobile manipulators, whose larger and more flexible reachable workspaces may allow them to present targets around the user that cannot be covered by a fixed-base manipulator. In principle, the same guided-reaching abstraction could also be extended to complementary interfaces, such as augmented-reality systems that visually indicate target locations [23, 24]. However, embodied robotic platforms retain important advantages for workplace deployment, including the ability to physically manipulate task objects, adapt the interaction geometry online, and integrate the estimated mobility profile directly into subsequent collaborative actions.

VI CONCLUSIONS

This work introduced the idea of enabling a robot to autonomously estimate the impairment coefficients that parameterise the goal-oriented functional ranges of motion of the upper limb of a user through handovers, with the future aim of tuning the interaction to the actual capabilities of the user and help towards an active recuperation process. Particularly, an active perception pipeline that estimates the impairment coefficients through a particle filter was described. The process was tested in simulation through a human model whose functional joint limits were parameterised by the same representation based on impairment coefficients. In real-life settings, the pipeline was validated by having a participant perform repetitive handovers with a robotic manipulator while wearing harnesses to prevent partially their movements within delimited ranges. The robot jointly estimated the impairment coefficients of the shoulder and elbow flexion with suitable accuracy with up to 25 handovers.

Although a robotic manipulator was used in this study, the approach is especially relevant to humanoid robots, whose human-like morphology, whole-body capabilities, and potential for natural face-to-face interaction make them particularly suitable for return-to-work scenarios. Unlike a fixed manipulator, a humanoid robot could adapt not only the handover trajectory but also its posture, positioning, and interaction strategy to the user’s functional limitations, enabling more personalised, socially engaging, and ecologically valid rehabilitation exercises.

ACKNOWLEDGMENTS

The authors acknowledge Mattia Leonori and Luca Fortini for the assistance with the Real-Move setup, and Luigi Pedata for the assistance with the Mobiwork robotic platform setup.

References

1. [1] L. Vianello, L. Penco, W. Gomes, Y. You, S. M. Anzalone, P. Maurice, V. Thomas, and S. Ivaldi, “Human-humanoid interaction and cooperation: a review,” Current Robotics Reports, vol. 2, no. 4, pp. 441–454, 2021.
2. [2] M. Lagomarsino, E. Merlo, A. Pupa, T. Birr, F. Krebs, C. Secchi, T. Asfour, and A. Ajoudani, “Intuitive programming, adaptive task planning, and dynamic role allocation in human–robot collaboration,” Annual Review of Control, Robotics, and Autonomous Systems, vol. 9, 2025.
3. [3] C. D. Rocha, I. Carneiro, M. Torres, H. P. Oliveira, E. Solteiro Pires, and M. F. Silva, “Post-stroke upper limb rehabilitation: clinical practices, compensatory movements, assessment, and trends,” Progress in Biomedical Engineering, vol. 7, no. 4, p. 042001, 2025.
4. [4] M. Lagomarsino, R. Arbaud, F. Tassi, and A. Ajoudani, “Mitigating compensatory movements in prosthesis users via adaptive collaborative robotics,” IEEE International Conference on Rehabilitation Robotics, pp. 470–475, 2025.
5. [5] S. Cai, P. Xie, G. Li, and L. Xie, “Compensation-corrective adaptive control strategy for upper-limb rehabilitation robots,” Robotics and Autonomous Systems, vol. 177, p. 104701, 2024.
6. [6] D. Torielli, L. Bertoni, L. Muratore, and N. Tsagarakis, “A laser-guided interaction interface for providing effective robot assistance to people with upper limbs impairments,” IEEE Robotics and Automation Letters, 2024.
7. [7] S. Poirier, F. Routhier, and A. Campeau-Lecours, “Voice control interface prototype for assistive robots for people living with upper limb disabilities,” in Proceedings of Int. Conf. on Rehabilitation Robotics. IEEE, 2019, pp. 46–52.
8. [8] P. Ardon, , M. E. Cabrera, E. Pairet, R. P. A. Petrick, R. P. A. Petrick, S. Ramamoorthy, K. S. Lohan, and M. Cakmak, “Affordance-aware handovers with human arm mobility constraints,” IEEE​ Robotics and Automation Letters, pp. ​3136–43, 2021.
9. [9] D. J. Gladstone, C. J. Danells, and S. E. Black, “The fugl-meyer assessment of motor recovery after stroke: a critical review of its measurement properties,” Neurorehabilitation and neural repair, vol. 16, no. 3, pp. 232–240, 2002.
10. [10] G. Chini, L. Fiori, A. Tatarelli, T. Varrecchia, F. Draicchio, and A. Ranavolo, “Indexes for motor performance assessment in job integration/reintegration of people with neuromuscular disorders: A systematic review,” Frontiers in Neurology, vol. 13, pp. 1–24, 9 2022.
11. [11] M. Lagomarsino and F. Tassi, “Towards online robot interaction adaptation to human upper-limb mobility impairments in return-to-work scenarios,” IEEE Robotics and Automation Letters, 2026.
12. [12] S. Keyvanian, M. J. Johnson, and N. Figueroa, “Learning realistic joint space boundaries for range of motion analysis of healthy and impaired human arms,” IEEE-RAS International Conference on Humanoid Robots, 2023.
13. [13] S. T. Lauman, L. J. Patton, P. Chen, S. Ravi, S. J. Kimatian, and S. E. Rebstock, “Assessing the accuracy and reliability of the monitored augmented rehabilitation system for measuring shoulder and elbow range of motion,” Sensors, vol. 25, no. 14, p. 4269, 2025.
14. [14] P. Khanna, T. Oppenheim, A. Tu-Chan, G. Abrams, and K. Ganguly, “Measuring arm and hand joint kinematics to estimate impairment during a functional reach and grasp task in stroke participants,” Neurorehabilitation and neural repair, vol. 37, p. 409, 6 2023.
15. [15] Z. Liu, Y. Ju, Y. Da, T. Silver, P. N. Thakkar, J. Li, J. Guo, K. Dimitropoulou, and T. Bhattacharjee, “Grace: Generalizing robot-assisted caregiving with user functionality embeddings,” ACM/IEEE International Conference on Human-Robot Interaction, pp. 686–695, 2025.
16. [16] D. M. A. Mehler, A. Reichenbach, J. Klein, and J. Diedrichsen, “Minimizing endpoint variability through reinforcement learning during reaching movements involving shoulder, elbow and wrist,” PloS one, vol. 12, no. 7, p. e0180803, 2017.
17. [17] A. Scano, E. Guanziroli, C. Brambilla, A. Specchia, L. M. Tosatti, and F. Molteni, “Kinematic instrumental assessment quantifies compensatory strategies in post-stroke patients,” Medical & Biological Engineering & Computing 2025 64:1, vol. 64, pp. 135–146, 9 2025.
18. [18] P. M. Djuric, J. H. Kotecha, J. Zhang, Y. Huang, T. Ghirmai, M. F. Bugallo, and J. Miguez, “Particle filtering,” IEEE signal processing magazine, vol. 20, no. 5, pp. 19–38, 2003.
19. [19] T. Grant, Y. Thomas, P. Gossman, and L. Berragan, “The use of simulation in occupational therapy education: A scoping review,” Australian Occupational Therapy Journal, vol. 68, no. 4, pp. 345–356, 2021.
20. [20] Safety of machinery — Anthropometric requirements for the design of workstations at machinery, International Organization for Standardization Std. ISO 14 738:2002, 2002.
21. [21] F. Semeraro, J. Leadbetter, and A. Cangelosi, “Trihrcbot: A robotic architecture for triadic human-robot collaboration through mediated object alignment,” in 2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 6703–6709.
22. [22] Real-Move, https://realmove.it/, 2026, accessed: 24 June 2026.
23. [23] S. G. Létourneau, H. Jin, E. Peters, R. Grewal, D. Ross, and C. Symonette, “Augmented reality-based finger joint range of motion measurement: assessment of reliability and concurrent validity,” The Journal of Hand Surgery, 2024.
24. [24] S. Basiratzadeh, E. D. Lemaire, and N. Baddour, “Augmented reality approach for marker-based posture measurement on smartphones,” in 2020 42nd annual international conference of the ieee engineering in medicine & biology society (EMBC). IEEE, 2020, pp. 4612–4615.