Yoga Personalization Challenge

• Yoga Personalization Problem is defined by dynamically selecting a subset of interdependent yoga items while ensuring safe, sequence-aware practice.
• It integrates AI for tailored recommendation, ubiquitous sensing for pose monitoring, and HCI for precise, real-time corrective feedback.
• Recent systems demonstrate effective real-time posture correction, personalized motion forecasting, and multimodal coaching using biofeedback.

The Yoga Personalization Problem is a decision-support problem concerned with helping a person choose, perform, and adapt yoga practices over time from a large, structured, and interdependent practice space. In the formulation centered on technology-assisted personalized yoga, yoga practices are executable items—primarily physical postures (āsana) and breathing routines (prāṇāyāma)—and personalization is not limited to recommending a few exercises once. It includes discovering a suitable subset of items, continuing practice over time while adjusting to changing abilities and near-term goals, sensing performance, detecting errors, recommending corrections, and adapting recommendations as health conditions and environment change (Kumar et al., 15 Aug 2025). Across recent systems, this problem appears in several forms: reference-pose-based posture correction from video, temporal modeling of pose execution, person-specific motion forecasting, multimodal adaptive coaching, and broader personalization of relaxation and contemplative experience (N, 2021, Gadhvi et al., 25 May 2025, Priisalu et al., 2023, Agarwal et al., 1 Apr 2026, Zhang et al., 2 Feb 2026).

1. Problem formulation as subset selection, monitoring, and adaptation

A central statement of the problem is that yoga personalization is not “recommend the top-$k$ poses.” It is the problem of recommending a subset of items to a user from a large yoga set and then monitoring whether those items are performed correctly. The practice space is a large interdependent set with sequence constraints, breath coordination, tradition-specific variants, and safety considerations. The paper that explicitly frames the topic identifies three core needs: discovering a suitable subset of yoga items, continuing practice over time while adjusting to changing abilities, interest, and near-term goals, and adapting to environmental and health changes, including injuries, chronic disease, and motor limitations (Kumar et al., 15 Aug 2025).

Within that framing, the main subproblems are item discovery or subset recommendation, pose sensing and recognition, correction recommendation, adaptation over time, data management, and interface design. The same source also separates the problem into three technical lenses: AI for recommending the right set of yoga items, Ubiquitous Computing (UC) for monitoring practice through sensing, and HCI for communicating feedback and recommendations effectively. This formulation is notable because it treats yoga not as a flat catalog of independent exercises but as a sequence-dependent, health-sensitive, and safety-constrained practice system (Kumar et al., 15 Aug 2025).

The distinction from generic recommendation is explicit. Yoga items are ordered embodied actions with physical execution correctness, breathing coordination, temporal dependencies, and evolving physiological relevance. A plausible implication is that standard recommender objectives based only on preference or engagement are insufficient unless they are coupled to sensing, correctness assessment, and adaptation.

2. Reference-pose comparison and posture-correction pipelines

One operationalization of yoga personalization is reference-pose-based posture correction. In Fitness tutor, a professional or trainer image provides a reference pose, a live webcam frame provides the user pose, and the system compares their body skeletons to decide whether the user should move arms or legs up, down, left, or right. The workflow is: take a reference image of the exercise or yoga pose; extract its PoseNet skeleton; store the reference skeleton as JSON; capture the live webcam feed; for each frame, extract the user skeleton; compare selected body-part coordinates or slopes; and generate feedback if the user deviates from the reference. The implementation uses the PoseNet model in TensorFlow, integrated with ml5.js, with visualization and skeleton handling done with p5.js. PoseNet identifies 17 body keypoints, but the system uses a subset including leftShoulder, rightShoulder, leftWrist, rightWrist, leftHip, rightHip, leftAnkle, and rightAnkle (N, 2021).

The comparison rule in that system is slope-based. For selected body-part pairs, the slope is

$\text{Slope} = \frac{y_2 - y_1}{x_2 - x_1}.$

The paper explicitly lists the comparisons Left shoulder → Left elbow, Right shoulder → Right elbow, Left hip → Left ankle, and Right hip → Right ankle. A user pose is accepted if the slope is within $\pm 0.5$ of the reference slope, i.e., if the slope difference is no larger than $0.5$. If one arm or leg is misaligned, the system displays a message identifying the part to correct. The result is directional, body-part-specific feedback rather than a purely binary decision (N, 2021).

A more biomechanical but still reference-driven formulation appears in the hybrid Edge-AI framework for real-time yoga pose detection and posture correction. There, input comes from a monocular RGB camera represented as $I_t \in R^{H \times W \times 3}$, followed by preprocessing $I_t' = f_{pre}(I_t)$. The framework names lightweight pose-estimation options including MoveNet, BlazePose, PoseNet, and MediaPipe. Detected pose keypoints are represented as

$P_t = \{(x_i, y_i, v_i)\}_{i=1}^{N}.$

The system computes biomechanical features such as joint angles, limb orientations, and body alignment parameters. For three points $A$, $B$, and $C$, the joint angle at $\text{Slope} = \frac{y_2 - y_1}{x_2 - x_1}.$0 is

$\text{Slope} = \frac{y_2 - y_1}{x_2 - x_1}.$1

and

$\text{Slope} = \frac{y_2 - y_1}{x_2 - x_1}.$2

These angles are compared with reference angles stored in a pose database through

$\text{Slope} = \frac{y_2 - y_1}{x_2 - x_1}.$3

The framework also defines an aggregate posture score,

$\text{Slope} = \frac{y_2 - y_1}{x_2 - x_1}.$4

with imperfect formatting in the paper, but the intended meaning is deviation-based penalization normalized by a maximum allowable deviation. If deviation exceeds a predefined threshold, the corresponding joint is marked incorrect and corrective feedback is generated through visual, text-based, and voice-based guidance (Haldar, 23 Mar 2026).

PosePilot extends this family of systems by making correction temporal and stage-wise rather than purely static. Its pipeline is: video capture of a yoga sequence; keypoint extraction using MediaPipe; angle computation from keypoints; pose recognition using an LSTM-based classifier; pose correction using a BiLSTM model with multi-head attention; error detection by comparing predicted or canonical motion against the user’s current angles; and personalized feedback shown in a GUI with a correction graph and highlighted errors. For correction it reduces the feature space from 680 angles to 9 key angles chosen with expert input: left shoulder, right shoulder, left elbow, right elbow, left hip, right hip, left knee, right knee, and neck. The error rule is motion-aware: if the difference between predicted and actual values exceeds 1.5 standard deviations, the point is flagged as an error, while values within one standard deviation of the predicted value are acceptable (Gadhvi et al., 25 May 2025).

3. Recognition backbones, datasets, and the identification layer

Pose recognition is the identification layer on which correction and recommendation are built. In the CLIP-based study of yoga pose analysis, the main benchmark is Yoga-82, with 82 leaf classes grouped into 20 superclasses and then 6 broad superclasses. After URL-based retrieval in spring 2023, 19.1K images were available for analysis. The study also constructs two six-pose subsets—Balasana, Dhanurasana, Marjaryasana, Sarvangasana, Ustrasana, and Utkatasana—to examine low-data and application-oriented settings. Zero-shot CLIP is reported as insufficient for fine-grained yoga recognition, but fine-tuning ViT-B/32 gives 98.8% accuracy on the unfiltered six-pose subset, 99.1% on the filtered six-pose subset, and 85.94% accuracy on the full 82-class setting, with precision = 86.1%, recall = 85.9%, and F1 = 85.7%. The reported inference time is about 7.1 ms per image, and the authors state that the model can be integrated into automated systems for posture evaluation and “a real-time personal yoga assistant for performance assessment” (Dobrzycki et al., 13 Jan 2025).

A separate benchmark analysis proposes SYD-Net, an end-to-end trainable model that adds patch-based attention (PbA) on top of a CNN backbone. The model decomposes the feature map into uniform and multi-scale patches, computes channel attention through self-attention across patches, computes spatial attention through GAP and GMP followed by an MLP, and combines them as

$\text{Slope} = \frac{y_2 - y_1}{x_2 - x_1}.$5

with final classification

$\text{Slope} = \frac{y_2 - y_1}{x_2 - x_1}.$6

On Yoga-82, the best reported result is 97.80% top-1 and 99.42% top-5 using Xception with patch set $\text{Slope} = \frac{y_2 - y_1}{x_2 - x_1}.$7 and ImageNet weights. On Yoga-107, the best reported result is 87.17% top-1 with DenseNet-201 and $\text{Slope} = \frac{y_2 - y_1}{x_2 - x_1}.$8. The paper is classification-focused rather than a full correction system, but it explicitly motivates yoga applications such as self-assessment, yoga assistance, self-coaching, and avoiding incorrect postures (Bera et al., 2023).

PosePilot contributes a distinct dataset for correction rather than only recognition. Its in-house video dataset contains 14 participants (10 male, 4 female), age range 17–25, 6 yoga asanas, four angles, 56 videos per pose, average duration about 7 seconds, and a total of 336 videos. On this dataset, the recognition model reports Average recognition accuracy: 97.52%, F1 score: 0.99, and 946 out of 970 frames correctly identified. The correction model reports Average MSE: 0.00138. For deployment, inference is optimized with TensorFlow Lite and INT8 quantization, where pose-recognition accuracy drops by only 1.0% and correction accuracy by 1.3%. Reported speed is 330.65 FPS for pose recognition and 6.42 FPS for pose correction (Gadhvi et al., 25 May 2025).

Taken together, these systems indicate that the recognition layer has progressed along at least three lines: fine-grained image classification, lightweight attention-enhanced classification, and sequence-aware edge correction. This suggests that the personalization bottleneck increasingly shifts from “what pose is this?” to “how should guidance change for this practitioner, at this moment, under these constraints?”

4. Temporal modeling and person-specific motion prediction

Temporal structure is central because yoga is not only the final posture. PosePilot makes this explicit: “the sequence of sub-poses leading into the asana matters.” The recognizer first reduces redundancy in variable-length videos by selecting key frames based on movement variability. For frame $\text{Slope} = \frac{y_2 - y_1}{x_2 - x_1}.$9, local mean and standard deviation over temporal window $\pm 0.5$0 are

$\pm 0.5$1

and

$\pm 0.5$2

with aggregated deviation

$\pm 0.5$3

Frames with local maxima in $\pm 0.5$4 are selected as key frames, and the best result is reported with $\pm 0.5$5 key frames. Correction is then handled by a two-layer BiLSTM with multi-head attention that models motion both forward and backward in time and forecasts the expected next joint-angle vector $\pm 0.5$6; deviations between predicted and actual joint states are used to identify errors (Gadhvi et al., 25 May 2025).

A more explicitly person-specific formulation appears in Personalized Pose Forecasting, which reformulates human pose forecasting so that the system predicts the motion of one particular individual rather than an average human. In the standard setup, a pose at time $\pm 0.5$7 for individual $\pm 0.5$8 is denoted $\pm 0.5$9, and the benchmark problem uses an observation window of length $0.5$0 to predict the next $0.5$1 poses. The paper evaluates predictions using Mean Per Joint Error (MPJE) and Mean Euler Angle (MEA) and notes that common benchmarks such as Human3.6M, HumanEva, and CMU Mocap do not reflect settings where a system observes the same person repeatedly over time (Priisalu et al., 2023).

The paper’s main contribution is a model-agnostic personalization method in which a neural network predictor $0.5$2 is corrected online by an autoregressive time-series model fit to the residuals or detrended sequence of that individual. For each joint and dimension, the residual dynamics are written as

$0.5$3

with forgetting-factor objective

$0.5$4

where $0.5$5 weights recent observations more heavily. The method is described as online, recursive, and low-cost, using roughly 50–500 parameters rather than the much larger parameter counts of neural baselines. The paper reports that AR correction improves early prediction steps and that AR models can be competitive with neural predictors in the personalized setting (Priisalu et al., 2023).

For yoga, the relevance is direct in the source text: instead of predicting an “average human” pose transition, the system should learn how one specific practitioner moves, bends, compensates, and transitions between poses over time. A plausible implication is that temporal personalization has at least two distinct meanings in this literature: stage-wise correction of the current movement trajectory, and online adaptation to an individual’s longer-term movement signature.

5. Multimodal adaptive coaching and broader personalization

FlexAI generalizes personalization beyond form correction by treating coaching as an online control problem rather than a static plan. The system uses two cameras, a smartwatch, and a wireless headset. The first camera is for full-body motion tracking and pose analysis via MediaPipe landmarks; the second camera is dedicated to facial expression analysis for pain and exertion cues. The smartwatch captures heart rate, heart rate variability, and step count, and the headset captures speech, breathing, and vocal cues of fatigue. At onboarding, the system creates a Physical Health Report using the WHO GPAQ, converts activity levels into MET categories, and collects height, weight, preferred intensity, previous injuries, and fitness goals. These signals are transformed into higher-level labels and fused into a JSON state representation for an LLM (Agarwal et al., 1 Apr 2026).

The inferencing pipeline is layered. Pain is estimated by segmenting the face with SAM and passing it to a fine-tuned ResNet-18 pain classifier. Fatigue is detected from voice-feature deviations from baseline, with deviations beyond an empirically determined threshold of 50–80% triggering a true fatigue state, and pilot testing with 5 users finding changes greater than 60% from baseline to be reliable for fatigue detection. Heart-rate state is based on the Karvonen Method,

$0.5$6

which is used to infer whether the user is Above, Target, or Below the desired zone. The reasoning module uses hierarchical prompting with an LLM, with separate prompt styles for inter-exercise transition decisions and intra-exercise interventions (Agarwal et al., 1 Apr 2026).

The yoga-specific portion of FlexAI is unusually explicit. It supports tree, warrior, and downward-facing dog; computes joint angles from MediaPipe landmarks; flags incorrect positioning; starts timing only when correct form is detected; pauses to provide corrective feedback when form deteriorates; offers midway and final interventions to sustain the pose; allows extension of the hold; and computes rest based on facial pain, fatigue signals, HR, and pose duration before moving to the next posture. Technical evaluation reports 79.3% accuracy for the pain classifier on the Delaware Pain Dataset and 76% in the test environment with self-reported pain as ground truth; repetition counting reaches 97.5% overall accuracy; and the visual feedback loop runs under 200 ms while the full audio feedback loop averages 1371.3 ms, with LLM inference at 485.2 ms and TTS generation at 784.5 ms. In a within-subject, counterbalanced study with 25 participants, FlexAI improved subjective experience relative to control, including enjoyment from 5.10 ± 1.81 to 6.33 ± 0.80 (p = 0.012), boredom from 2.95 ± 2.01 to 1.62 ± 0.80 (p = 0.0096), accomplishment from 4.57 ± 1.80 to 5.52 ± 1.12 (p = 0.0445), and frustration from 2.52 ± 2.14 to 1.43 ± 0.81 (p = 0.0311) (Agarwal et al., 1 Apr 2026).

Broader work on relaxation and contemplative personalization extends the problem beyond pose geometry. ASafePlace addresses the challenge of making relaxation, meditation, or contemplative practice feel deeply personal rather than generic. Its workflow begins with guided imagination of a safe place, followed by a sketch, therapeutic dialogue, AI-assisted co-creation of a personalized 360° environment, personalized audio guidance generation, and a relaxation phase in VR where biofeedback appears as particles moving with breathing. The visual pipeline uses Stable Diffusion web UI, Inazuma V10, Latent Labs 360 LoRA, ControlNet, and LCM LoRA; the audio guidance uses OpenAI GPT-4o and GPT-SoVITS TTS. In a 2×2 between-subjects factorial design with N = 52, the combination of personalization and biofeedback produced the largest improvement in HRV-SDNN (+42.5% ± 20.1%) and the largest reduction in respiration rate (-45.2% ± 35.5%), with significant interaction effects for both measures. The paper interprets the mechanism in terms of familiarity, presence, symbolism, autobiographical memory, and emotional connection (Zhang et al., 2 Feb 2026).

This broader literature suggests that yoga personalization may need to incorporate not only alignment and fatigue, but also breath, affect, sensory preferences, and personally resonant guidance. That inference is directly encouraged by the source texts on adaptive coaching and memory-based relaxation personalization.

6. Limitations, unresolved questions, and research outlook

Several limitations recur across the literature. The earliest pose-correction systems remain constrained by 2D image skeletons, camera angle sensitivity, lighting problems, and the fact that some poses are not well identified, so no useful correction can be given. In that line of work, future improvement is explicitly linked to 3D pose estimation (N, 2021). The hybrid Edge-AI framework for real-time yoga correction similarly states that it is primarily a proposed framework: it specifies datasets, tools, training setup, and intended evaluation metrics, but does not report actual numeric experimental results for the proposed system. It also notes that the current design does not implement body-type adaptation, flexibility-aware calibration, a user profile, or adaptive thresholds per user (Haldar, 23 Mar 2026).

FlexAI demonstrates richer adaptation but also lists important caveats. The control condition is minimal and non-adaptive; the study is short-term and single-session; long-term adherence, learning, and injury outcomes are unknown; the system relies on correlational inference rather than medical diagnosis; audio latency is a real issue for time-sensitive cues; and the participant pool is limited to young adults. The authors therefore recommend conservative adaptation, user-specific baselines, and strong safety guardrails (Agarwal et al., 1 Apr 2026). Personalized Pose Forecasting likewise presents its method as a first step: AR models have exponentially decreasing prediction power and are mainly useful for short-term correction, while longer-range personalization may require meta-learning or key-pose-based methods (Priisalu et al., 2023).

At the level of the field, open questions remain about representation, ground truth, data fusion, and evaluation. The vision paper on technology-assisted personalized yoga identifies obscure Sanskrit names, inconsistent naming across traditions, multimodal synchronization, storage and retrieval at scale, sequence representation, and the distinction between pose error and transition error as unresolved issues. Its Surya Namaskar case study shows why the problem is difficult: the sequence has 12 choreographed poses, is often described as 24 steps, contains breath and mantra coordination, and varies across traditions in both naming and execution. The same paper proposes future directions including AI/ML-based analysis, knowledge graphs, live pose recognition and real-time feedback, group recommendation methods, sensor subset optimization, and privacy-preserving and secure infrastructure (Kumar et al., 15 Aug 2025).

The resulting picture is not of a single algorithmic task but of a layered research agenda. Yoga personalization includes subset selection over a constrained practice graph, recognition of what is being performed, evaluation of how it is being performed, real-time correction, longitudinal adaptation to the same individual, and safe interaction design. The most developed components currently appear to be fine-grained recognition, posture correction, and multimodal intervention. The least settled components are long-term regimen recommendation, robust personalization across traditions and body types, and evaluation of safety, adherence, and benefit over extended periods.