Learning Varying Physical Therapist-Patient Interactions for Robot-mediated Upper Limb Task-Specific Training

Technical Summary: Learning Varying Physical Therapist-Patient Interactions for Robot-mediated Upper Limb Task-Specific Training

Problem Statement
Upper extremity motor recovery is positively linked to Task-Specific Training (TST) and sufficient therapy dosage. While rehabilitation robots can increase dosage through controlled, repetitive treatment, they have not yet demonstrated significant benefits over conventional therapy. This limitation is attributed to two main factors: the inaccurate robotic representation of personalized physical therapist-patient interactions and a lack of practice variability during TST. Current robotic interventions often rely on predefined physical Human-Robot Interaction (pHRI) schemes (e.g., assist-as-needed, resist) that fail to capture the nuanced, personalized forces therapists apply to patients with varying neurological impairments. Furthermore, existing Learning-from-Demonstration (LfD) approaches for Robot-mediated physical Human-Human Interaction (RHHI) often focus on reproducing interactions only within previously seen conditions, lacking the ability to generalize to new task variations (e.g., different positions, speeds, or objects) without constant therapist redemonstration.

Methodology
The authors propose a framework to learn and generalize personalized therapist-patient interactions across varying task conditions using a small number of demonstrations.

1. Interaction Modeling: The physical interaction is defined as a vector $\gamma_t$ comprising patient joint positions ($q_t$), velocities ($\dot{q}_t$), and therapist-applied joint torques ($\tau_t$). The goal is to learn a mapping $g_{a,e}$ from patient kinematics to therapist torques for a specific therapist-patient pair ($a$) and task ($e$).
2. Proposed Framework (TPGMM): The core contribution is the use of Task-Parameterised Gaussian Mixture Models (TPGMM).
   • Frame Transformation: Demonstrations are transformed into multiple task-dependent reference frames (e.g., start, via, and end points of a task). In this study, these frames are defined by specific joint kinematic configurations (joint positions and velocities) rather than Cartesian space.
   • Learning: A Gaussian Mixture Model (GMM) is learned in each reference frame using the Expectation-Maximization (EM) algorithm. The models are combined using a product of Gaussians to form a reduced GMM.
   • Regression: For new task variations, the model conditions the learned GMMs on new reference frames defined by the new task points. Gaussian Mixture Regression (GMR) is then used to estimate the therapist-applied torques based on the patient's current joint kinematics.
3. Benchmark (Look-Up Table): The TPGMM approach is compared against a Look-Up Table (LUT) method (based on Luciani et al.), which retrieves the nearest neighbor demonstration based on Euclidean distance in joint kinematic space and applies a low-pass filter to the resulting torque.
4. Experimental Setup:
   • Participants: 14 pairs of physiotherapy students (acting as therapist and patient) were recruited.
   • Tasks: Three task-specific exercises (reaching for a cup, using utensils, pouring water) were performed. Each task had three patient personas with varying impairment levels (FMA-UE scores 8, 32, and 56) and six variations involving changes in distance, orientation, or volume.
   • Data Collection: A wearable sensor system (XSENS motion capture and three 6-DOF force-torque sensors on shoulder, upper arm, and forearm) recorded joint kinematics and interaction wrenches.
   • Evaluation: The models were trained on four variations and tested on two unseen variations (generalization) and held-out trials from training variations (validation). Performance was measured using normalized average reconstruction error ($\epsilon_n$), normalized peak torque difference ($\Delta\tau_{peak}$), and coverage ($C_n$).

Key Contributions

• Generalization of Interactions: The study demonstrates a method to learn therapist-patient interaction profiles from few demonstrations and generalize them to unseen task variations (different task points and conditions), addressing the gap in current RHHI literature which often only reproduces seen conditions.
• TPGMM for Torque Prediction: The application of TPGMM specifically for mapping patient joint kinematics to therapist-applied joint torques in a high-dimensional joint space, rather than just end-effector positions.
• Empirical Comparison: A rigorous comparison between a statistical generalization model (TPGMM) and a direct retrieval method (LUT) in the context of upper limb rehabilitation with varying patient personas and task complexities.

Results

• Validation (Seen Variations): Both TPGMM and LUT successfully reconstructed interactions within the therapist's behavioral variability. The LUT method showed slightly better fidelity in capturing fine details (peak torques) compared to TPGMM.
• Generalization (Unseen Variations): Both methods produced reconstructions that deviated slightly from the actual therapist interaction but remained close to the intended behavior boundaries.
  • TPGMM Performance: TPGMM demonstrated a small but statistically significant advantage over LUT in terms of reconstruction error ($\epsilon_n$) and coverage ($C_n$) for unseen variations.
  • Task Complexity Effect: A counter-intuitive finding was observed regarding task complexity. In the validation case (seen variations), increased task complexity (from Persona 1 to 3) degraded performance for both methods. However, in the generalization case (unseen variations), increased task complexity significantly improved performance (lower error, higher coverage) for both methods.
• Limitations: Neither method perfectly captured peak torque details ($\Delta\tau_{peak}$ remained high), indicating a trade-off between generalization capability and fine-grained interaction fidelity.

Significance and Claims
The paper claims that the proposed TPGMM framework offers a viable approach for robot-mediated TST that preserves personalized therapist-patient interactions across varying practice conditions. By learning the "latent features" of physical interactions, the system reduces the need for therapists to constantly redemonstrate interactions for every new task variation.

The authors maintain a modest tone regarding clinical applicability. They acknowledge that the study was conducted with physiotherapy students acting as patients and therapists, using a constrained sensor setup (three contact points) that differs from real-world clinical interactions. The framework was evaluated offline on unconstrained kinematic data. Consequently, the paper does not claim immediate clinical efficacy but suggests the proposed methods are suitable for such scenarios and warrants further integration with actual rehabilitation robotic systems and pilot studies with stroke patients to validate real-world efficacy. The work serves as a preliminary step toward enabling robots to adapt to the dynamic and personalized nature of human therapy.