TherapyGym: Tech-Mediated Therapeutic Environments

• TherapyGym is a cross-domain research paradigm that integrates sensor technology, VR, wearables, and robotics to facilitate both physical rehabilitation and mental-health support.
• It employs adaptive training, procedural content generation, and structured feedback to tailor therapy tasks and monitor performance with measurable outcomes.
• Systems under TherapyGym achieve high-precision measurement and dynamic feedback, enhancing motor rehabilitation, remote therapy, and clinical chatbot evaluation.

TherapyGym denotes a research paradigm of technology-mediated therapeutic environments in which sensing, adaptive software, and structured feedback are used to support rehabilitation, therapeutic communication, and longitudinal monitoring. In the physical-rehabilitation literature synthesized under this label, TherapyGym spans marker-less motion capture, VR and mobile VR exergames, wearable and physiological sensing, procedural content generation, robotic targets, and exoskeleton-mediated gait training for upper-limb, gait, balance, neck-posture, low-back-pain, and daily-function tasks (Dhingra et al., 2023, Stanica et al., 2020, Küçüktabak et al., 21 Jul 2025). In a distinct mental-health usage, “TherapyGym” is also the name of a framework for evaluating and aligning therapy chatbots according to psychotherapy fidelity and safety rather than generic dialogue quality (Huang et al., 23 Feb 2026). This suggests that TherapyGym is best understood as a cross-domain therapeutic design space rather than a single fixed product.

1. Conceptual scope

Across the cited work, TherapyGym encompasses several technically different but structurally related systems: pediatric and adult exergames, immersive neurorehabilitation platforms, webcam-based telerehabilitation, older-adult virtual gyms, therapist-guided robotic exercise, and clinically grounded conversational systems. Upper-limb exemplars include “Neurorehab: An Interface for Rehabilitation” (Dhingra et al., 2023), “Jcave: A 3D Interactive Game to Assist Home Physiotherapy Rehabilitation” (Elrefaei et al., 2019), and “Virtual Therapy Exergame for Upper Extremity Rehabilitation Using Smart Wearable Sensors” (Baron et al., 2023). Immersive and remote systems include “Flexible Virtual Reality System for Neurorehabilitation and Quality of Life Improvement” (Stanica et al., 2020), “TOSHFA: A Mobile VR-Based System for Pose-Guided Exercise Rehabilitation for Low Back Pain” (Mohamed et al., 24 Jan 2026), and “Necknasium: A Virtual Reality Rehabilitation Game for Managing Faulty Neck Posture” (Youssef et al., 2023). Social and embodied forms include the tablet-based virtual gym for older adults (Baez et al., 2016), the mobile-manipulator system “Stretch with Stretch” (Lamsey et al., 2023), and therapist–patient exoskeleton coupling for gait therapy (Küçüktabak et al., 21 Jul 2025). In mental-health computing, the lineage also includes screen-mediated CBT support such as gNats Island and MindBalance (Coyle et al., 2013), as well as the therapy-chatbot evaluation framework “TherapyGym: Evaluating and Aligning Clinical Fidelity and Safety in Therapy Chatbots” (Huang et al., 23 Feb 2026).

Domain	Representative systems	Primary modality
Motor rehabilitation	Neurorehab; JCave; VR exergame with CNT sleeve	Kinect, VR, wearable sensing
Immersive and remote rehabilitation	INREX-VR; TOSHFA; Necknasium	VR, webcam pose estimation, HMD tracking
Social and embodied therapy	Gymcentral; Stretch with Stretch; TEPI	online group exercise, mobile manipulator, exoskeleton coupling
Mental-health support and evaluation	gNats Island; MindBalance; TherapyGym	game-mediated CBT, online support, LLM evaluation

The scope is not restricted to screen-based interfaces. “Harnessing the ‘Reactive Falling Effect’ for rehabilitation and performance boosting” defines Logic Workout as a training method based on rolling, deformation, and spring instability on small fitballs, with a cohort of 18 participants and self-reported rehabilitation and performance outcomes (Sornette et al., 16 Jun 2025). A plausible implication is that TherapyGym, in the broader sense used here, includes both digital and physically instrumented environments so long as they preserve three recurrent properties: therapeutic task structure, measurable performance, and adaptive progression.

2. Sensing, measurement, and computational substrate

A central feature of TherapyGym systems is the conversion of bodily activity into computable state. Kinect-based systems remain important in this lineage. Neurorehab uses Microsoft Kinect v2.0 with Unity 5 3D, C#, Kinect Free SDK, Raw Mocap Data for Mecanim, MS-Kinect API, and Final IK; it tracks targets between 0.5 and 5 meters at approximately 30 frames per second, with random depth error of a few millimeters at 0.5 m and up to approximately 4 cm at 5 m (Dhingra et al., 2023). A different server-assisted pipeline is implemented in the HRNet rehabilitation monitoring system: an iOS app uploads videos, HRNet-W32 estimates 17 2D keypoints per frame, erroneous coordinates are corrected via Z-score outlier detection and temporal correlation checks, similarity is computed from KL divergence over 11 joint-angle time series, and repetition counting uses Savitzky–Golay smoothing with wavelet-based peak detection; the best similarity-calculation result occurs at $T = 0.2$, where Precision = 0.931, Recall = 0.900, and $F1 \approx 0.915$, while soft repetition-count accuracy exceeds 90% across all six tested actions (Hung et al., 2023). TOSHFA replaces depth sensing with a 720p webcam, MediaPipe Pose, and low-latency UDP streaming to a smartphone in a cardboard-style VR headset; it tracks 33 skeletal landmarks, runs pose inference typically under 20 ms, uses 15–25% CPU, and keeps one-way UDP latency under 10 ms with end-to-end latency below 100 ms (Mohamed et al., 24 Jan 2026).

Wearable and multimodal systems expand this measurement layer beyond pose alone. INREX-VR combines an HTC Vive Cosmos Elite, controllers, Vive trackers, a Myo armband streaming EMG at 200 Hz, and a Mi Fit 3 heart-rate bracelet; joint mobility is computed from IK and vector-angle calculations, while sessions and biometrics are synchronized through Firebase-backed telemedicine workflows (Stanica et al., 2020). The telerehabilitation exergame thesis “Exergames for telerehabilitation” uses five BNO055 MARG sensors on the upper limbs and back for triplanar ROM assessment, RehabFork for eating-task rehabilitation, and a load-cell-based grasp rehabilitator for pincer and lifting tasks (Bethi, 2020). The CNT-sleeve VR exergame adds joint-level sensing to controller tracking by placing a flexible carbon nanotube sleeve over the dominant elbow and computing normalized resistance change as

$$\mathrm{ResistanceChange}(\%) = \frac{(R - R_m)\times 100}{R_m},$$

where $R$ is resistance during stretch and $R_m$ is the minimum resistance at no stretch; negative resistance changes are discarded as outliers, and no angle calibration is reported (Baron et al., 2023).

Several systems transform raw motion into clinically interpretable feature spaces. UbiPhysio uses either 17 IMUs at 60 Hz or vision-based 3D pose, normalizes motion to a 24-joint skeleton, extracts 287 low-level features and 28 clinically designed biomechanical features per timestep, and classifies action type with a 1D ConvNet that achieves macro $F1 \approx 0.9636 \pm 0.0067$ (Wang et al., 2023). This emphasis on clinically curated feature engineering, rather than only endpoint tracking, recurs throughout TherapyGym systems and underwrites later stages of feedback and adaptation.

3. Adaptive training, procedural generation, and feedback

Adaptation in TherapyGym ranges from simple rule-based difficulty changes to explicit optimization and LLM-mediated intervention. Neurorehab implements dynamic difficulty by changing target size: as the user succeeds, the goal reduces in size, and if the user misses repeatedly, the goal increases in size (Dhingra et al., 2023). Necknasium organizes cervical retraction into six levels: Levels 1–3 require 30 retractions at 30%, 60%, and 90% of maximum retraction, while Levels 4–6 retain those ROM fractions and increase time-under-tension; in the prototype evaluation, Level 3 used movements lasting at least 6 seconds and Level 6 used at least 10 seconds (Youssef et al., 2023). “Stretch with Stretch” formalizes adaptation through exercise-specific target geometry. For the seated forward kick, the anchor point and difficulty direction are defined as

$$X_{0,\mathrm{kick}} = [l_U + l_L \sin \theta_K,\ 0.5 w_H,\ l_L \cos \theta_K], \qquad \bar d_{\mathrm{kick}} = [0.15,\ 0,\ 0.50],$$

with

$$X_{\min} = X_0 - 0.5\bar d,\quad X_{\max} = X_0 + 0.5\bar d,\quad X_{\mathrm{target}} = X_{\min} + F_{\mathrm{diff}}\bar d,$$

and difficulty updated by $\Delta F_{\mathrm{diff}} = +\delta, 0, -\delta$ for Excellent, Good, and Poor performance, with $\delta = 0.2$ (Lamsey et al., 2023).

A more formal adaptive framework is developed in “Intelligent Physiotherapy Through Procedural Content Generation.” There, procedural content generation and player modelling define content $F1 \approx 0.915$0 from a patient state $F1 \approx 0.915$1, optimize a utility $F1 \approx 0.915$2 under safety constraints, and update difficulty according to

$F1 \approx 0.915$3

with therapist goals and safety-critical movement envelopes constraining the search space (Esfahlani et al., 2018). “Using Learnable Physics for Real-Time Exercise Form Recommendations” pushes this further by combining MediaPipe pose recognition, peak-prominence repetition counting, and an Interaction Network that predicts expected motion trajectories; diagnosis operates on residual time series transformed into frequency-domain error signatures, yielding weighted $F1 \approx 0.915$4 scores of 0.94 for squats, 0.98 for push-ups, 0.97 for lunges, 0.98 for sit-ups, 0.98 for shoulder press, and 0.88 for front raise, with mean lag of 0.55 s for squats, 0.39 s for sit-ups, 0.36 s for push-ups, and 0.54 s for lunges (Jaiswal et al., 2023).

Feedback generation increasingly uses natural language rather than only scores, timers, or binary success cues. UbiPhysio first tokenizes action sequences with a VQ-VAE, then generates fine-grained movement descriptions and retrieval-enhanced physiotherapy feedback using action type, movement-pattern text, and a physio-curated knowledge base of low-demand and high-demand strategies (Wang et al., 2023). FlexAI performs label-level multimodal fusion of form error, pain class, fatigue state, and heart-rate zone, passes the resulting JSON into hierarchical LLM prompts, constrains inter-exercise rest planning to a maximum of 60 seconds, and constrains intra-exercise interventions to at most 15 words (Agarwal et al., 1 Apr 2026). A plausible implication is that TherapyGym systems are moving from kinematic correctness checking toward closed-loop coaching in which measurement, risk assessment, and language generation are all part of the control architecture.

4. Therapist mediation, sociality, and embodied interaction

A persistent theme in the TherapyGym literature is that technology is not primarily used to eliminate the therapist, but to reconfigure the therapeutic relationship. In the mental-health systems discussed by Coyle and Doherty, gNats Island places therapist and adolescent together at one computer, using island exploration, metaphors such as “catching, trapping and swatting gNats,” and screen-mediated dialogue to reduce direct eye-contact pressure and support client-centered CBT conversation; MindBalance uses asynchronous supporter feedback on shared CBT exercises, personal profiles, and thoughts–feelings–behaviours charts to create “some degree of social presence” in online depression treatment (Coyle et al., 2013). In both cases, the therapist role shifts from direct interrogator to collaborator or supporter, while software acts as a mediating object.

The same logic appears in physical rehabilitation. The tablet-based Gymcentral virtual gym for older adults organizes therapy around a Reception, Locker Room, Classroom, Progress Report, Training Schedule, public bulletin board, and private messages; participants receive individualized OTAGO-based exercises while training “together” through avatars, and differences in individual levels are hidden from others (Baez et al., 2016). In the pilot trial, persistence was 85.4% ± 16.1 in the Social group versus 64.2% ± 24.1 in the Control group, and the Group effect on persistence was significant, $F1 \approx 0.915$5, while the interaction Group × Initial Leg Muscle Strength was also significant, $F1 \approx 0.915$6, indicating that initial fitness affected adherence in individual training but not in the virtual-gym social condition (Baez et al., 2016).

Embodied therapist presence is made literal in recent robotic and software-generation systems. “Stretch with Stretch” uses a Hello Robot Stretch RE1 mobile manipulator and a soft-bubble end effector to present personalized physical targets for seated and standing exercises, with verbal prompts, sound effects, and cognitive dual-tasking (Lamsey et al., 2023). “Therapist-Exoskeleton-Patient Interaction: An Immersive Gait Therapy” couples therapist and patient through two lower-limb exoskeletons, mirror-mapped at the hips and knees through virtual spring–damper elements, so that the therapist both guides motion and feels patient deficits as interaction torques (Küçüktabak et al., 21 Jul 2025). “Clinician-Directed LLM Software Generation for Therapeutic Interventions in Physical Rehabilitation” preserves clinician authorship in a different way: 20 licensed physical and occupational therapists created 40 individualized upper-extremity programs comprising 398 instructions, and the LLM translated those prescriptions into executable software rather than selecting from a fixed exercise library (Kim et al., 23 Nov 2025). Across these systems, human expertise remains central, but its embodiment shifts among co-located guidance, social co-presence, robotic mediation, and program synthesis.

5. Fidelity and safety in conversational TherapyGym

In the therapy-chatbot literature, TherapyGym has a narrower and explicitly psychotherapeutic meaning. “TherapyGym: Evaluating and Aligning Clinical Fidelity and Safety in Therapy Chatbots” argues that metrics such as BLEU, ROUGE, MT-Bench, and HELM fail to capture the clinically critical dimensions of psychotherapy, and replaces them with two pillars: fidelity and safety (Huang et al., 23 Feb 2026). Fidelity is operationalized with the Beck Institute’s Cognitive Therapy Rating Scale (CTRS), an 11-aspect session-level instrument covering Agenda, Feedback, Understanding, Interpersonal Effectiveness, Collaboration, Pacing and Efficient Use of Time, Guided Discovery, Focusing on Key Cognitions or Behaviors, Strategy for Change, Application of CBT Techniques, and Homework, each scored from 0 to 6 (Huang et al., 23 Feb 2026). Safety is modeled with four binary session-level categories: provide medical opinion/medication; fail to address crisis and imminent risk; fail to address abuse; and failure to address functional impairment (Huang et al., 23 Feb 2026).

The framework includes TherapyJudgeBench, a validation set of 116 multi-turn CBT-style dialogues with 1,270 expert ratings. Human–human reliability on CTRS items averages approximately Krippendorff’s $F1 \approx 0.915$7, median 0.55, with average Spearman $F1 \approx 0.915$8; the best LLM judge configuration achieves average Spearman $F1 \approx 0.915$9 across CTRS skills and 99% safety accuracy against expert annotations (Huang et al., 23 Feb 2026). TherapyGym then turns these evaluators into reward signals for reinforcement learning with simulated patients derived from CBT cognitive models. CTRS scores are normalized as

$$\mathrm{ResistanceChange}(\%) = \frac{(R - R_m)\times 100}{R_m},$$0

and combined with safety penalties in a composite reward $$\mathrm{ResistanceChange}(\%) = \frac{(R - R_m)\times 100}{R_m},$$1 over a reliable subset $$\mathrm{ResistanceChange}(\%) = \frac{(R - R_m)\times 100}{R_m},$$2 of CTRS skills (Huang et al., 23 Feb 2026).

The reported alignment results are substantial. Under blinded human ratings, average CTRS for Qwen3-4B increases from 0.10 to 0.60 after GRPO training; under the LLM judge, it rises from 0.16 to 0.59. Dialogue-level safety violations under human ratings decrease from 0.38 to 0.20, and removing the safety penalty raises the violation rate under the LLM judge from 0.13 to 0.43 while slightly lowering average CTRS from 0.59 to 0.53 (Huang et al., 23 Feb 2026). This makes TherapyGym one of the few systems in the broader corpus where evaluation is explicitly anchored in a validated psychotherapy instrument rather than generic conversational quality.

6. Empirical status, controversies, and open problems

The empirical record for TherapyGym is strongest in instrumentation, feasibility, and short-term engagement, but heterogeneous in clinical maturity. INREX-VR reports average joint-angle evaluation accuracy of 95.59% against goniometer measurements (Stanica et al., 2020). The HRNet monitoring system reports $$\mathrm{ResistanceChange}(\%) = \frac{(R - R_m)\times 100}{R_m},$$3 for similarity calculation and soft repetition-count accuracy above 90% (Hung et al., 2023). The CNT-sleeve VR exergame finds a significant Orientation × Configuration interaction for normalized task completion time, $$\mathrm{ResistanceChange}(\%) = \frac{(R - R_m)\times 100}{R_m},$$4, and significant effects of orientation, configuration, and their interaction on normalized resistance change, but no significant differences in normalized mistakes (Baron et al., 2023). TOSHFA’s pilot study with 20 participants reports mean SUS 47.4, indicating marginal usability, while GEQ results show high positive affect and enjoyment despite technical friction (Mohamed et al., 24 Jan 2026).

Evidence for adherence and user acceptance is stronger than evidence for long-term clinical outcome in many systems. In the older-adult virtual-gym trial, both groups improved on Chair-stand and Timed Up & Go, and the social virtual-gym condition attenuated the usual relationship between initial fitness and adherence (Baez et al., 2016). In “Stretch with Stretch,” 10 participants with Parkinson’s disease rated perceived usefulness, ease of use, and attitude significantly above neutral, with mild to moderate perceived exertion (Lamsey et al., 2023). FlexAI’s controlled study with 25 participants shows significantly greater enjoyment, stronger sense of accomplishment, and significantly lower boredom and frustration than a static control system; trainer review rates the LLM interventions as safe ($$\mathrm{ResistanceChange}(\%) = \frac{(R - R_m)\times 100}{R_m},$$5), appropriate ($$\mathrm{ResistanceChange}(\%) = \frac{(R - R_m)\times 100}{R_m},$$6), and timely ($$\mathrm{ResistanceChange}(\%) = \frac{(R - R_m)\times 100}{R_m},$$7) (Agarwal et al., 1 Apr 2026). In clinician-directed software generation, the LLM implements 100% of the 40 personalized prescriptions compared with 55% for the template benchmark, correctly delivers 397/398 instructions as prescribed, monitors performance with 88.44% accuracy, and is judged safe by 90% of therapists, with 75% expressing willingness to adopt it (Kim et al., 23 Nov 2025).

At the same time, the literature contains substantial methodological asymmetry. TEPI gait therapy reports significant gains over conventional therapist-guided treadmill walking in ankle workspace area, step length, step height, and selected EMG channels in eight chronic stroke patients (Küçüktabak et al., 21 Jul 2025). Logic Workout reports complete pain resolution in several chronic cases and performance gains in a cohort of 18 participants, but outcomes are preliminary and self-reported, with no effect sizes, confidence intervals, or $$\mathrm{ResistanceChange}(\%) = \frac{(R - R_m)\times 100}{R_m},$$8-values (Sornette et al., 16 Jun 2025). Several other systems explicitly state that clinical evaluation outcomes, statistical analyses, calibration protocols, or detailed signal-processing pipelines are not reported, including Neurorehab and the PCG-based intelligent physiotherapy prototype (Dhingra et al., 2023, Esfahlani et al., 2018). This suggests that TherapyGym research, taken as a whole, is currently more mature as a systems-engineering and human–computer-interaction enterprise than as a uniformly validated clinical intervention science. The major open problems are therefore not only better sensing or richer feedback, but also standardized outcome reporting, larger controlled trials, clearer safety governance, and stronger integration between adaptive computation and clinician accountability.