WristPP: Wrist-Worn Hand Pose & Pressure Sensing

• WristPP is a wrist-worn system that uses a palmar-side 180° fisheye camera to simultaneously capture 3D hand pose and per-vertex pressure in real time.
• It employs a multi-stage annotation pipeline with a ViT backbone and Hand-VQ-VAE to recover detailed hand meshes and pressure maps from a single RGB frame.
• Performance metrics, including an MPJPE of 3.0 mm and contact accuracy above 97%, demonstrate its superiority over head-mounted baselines for mobile interaction.

Searching arXiv for the exact topic and closely related wrist papers to ground the article in current literature. Search query: WristPP arXiv (Xi et al., 28 Feb 2026) hand pose pressure estimation WristPP denotes a wrist-worn sensing paradigm centered on relocating rich hand-state inference to the wrist, with the most specific instantiation being the camera-based system "WristPP: A Wrist-Worn System for Hand Pose And Pressure Estimation" (Xi et al., 28 Feb 2026). In that system, a palmar-side fisheye camera continuously observes the hand and nearby contact region, enabling simultaneous estimation of 3D hand pose and per-vertex pressure from a single wide-FOV RGB frame in real time. Across adjacent literature, the same naming pattern also appears in discussions of wrist photoplethysmography, wrist haptics, and wrist-centered prosthetic or robotic control, all of which share a common systems objective: shifting sensing or feedback away from the fingertips or external infrastructure while preserving task-relevant information (Tarniceriu et al., 2017, Sarac et al., 2022, Sarac et al., 2022, Vasile et al., 24 Feb 2025, Xi et al., 28 Feb 2026).

1. Conceptual scope

In its narrow sense, WristPP is the 2026 wrist-worn visual system for joint hand-pose and pressure inference (Xi et al., 28 Feb 2026). Its stated target is “accurate 3D hand pose and pressure sensing” in “mobile scenarios,” using a wrist-mounted, palmar-side camera rather than head-mounted vision or surface instrumentation. The system estimates both a full 3D hand mesh and dense pressure over mesh vertices, and is evaluated not only as an offline prediction model but also as an interaction substrate for pointing, virtual touchpad use, and large-display control (Xi et al., 28 Feb 2026).

A broader reading is also justified by nearby work. Wrist-centered haptic studies investigate whether contact and mechanical properties normally perceived at the fingertips can be retargeted to the wrist without destroying task performance (Sarac et al., 2022, Sarac et al., 2022). Wrist photoplethysmography studies ask whether inter-beat intervals or identity-specific cardiovascular signatures can be extracted unobtrusively from wrist-worn optical sensing (Tarniceriu et al., 2017, Shao et al., 19 Aug 2025). Prosthetic and robotic wrist papers treat the wrist as a critical proximal degree of freedom for shared autonomy, variable stiffness, or manipulation efficiency (Milazzo et al., 2023, Vasile et al., 24 Feb 2025, Sulaiman et al., 13 Jan 2026, Merritt et al., 1 Apr 2026). This suggests a recurring research thesis: the wrist is not merely a mounting site, but a computationally productive sensing and actuation locus.

2. WristPP as a camera-based pose-and-pressure system

The defining WristPP hardware places a 180° FOV ultra-wide RGB camera on the palmar side of the wrist, paired with a Raspberry Pi Zero 2 W, a 2000 mAh Li-Po battery, and a custom 3D-printed enclosure with a magnetized 90° fold-out hinge (Xi et al., 28 Feb 2026). The hinge supports a stowed 0° position and a working 90° position. Reported hardware cost is about USD 50. In the real-world study, the wireless configuration transmits 512×512 at 30 fps over 2.4 GHz WLAN and runs for about 3 hours at roughly 1.5 W (Xi et al., 28 Feb 2026).

The central systems claim is that this wrist perspective resolves a gap left by existing approaches. Head-mounted systems require the hand to remain in the headset frustum and can induce “gorilla-arm” fatigue, while detailed pressure inference is difficult from distant viewpoints because pressure cues are subtle and sparse. WristPP addresses this by anchoring the sensor to the body and placing it close to the fingers and contact region, thereby supporting both mid-air gestures and pressure-aware interaction on ordinary near-planar surfaces (Xi et al., 28 Feb 2026).

The collected dataset contains 133,000 frames from 20 participants, with 48 on-plane gestures and 28 mid-air gestures (Xi et al., 28 Feb 2026). The on-plane set spans actions such as single-finger and multi-finger presses, sliding, pinching, rolling presses, circular motion, palm-side and palm-center presses, and light-touch and clench variants. The mid-air set includes 10 ASL letters, fist, OK sign, finger heart, grasp, pinch, pointing, rock gesture, thumbs-up, and claw pose. A subset of 5 participants was also recorded with a head-mounted RGB camera for baseline comparison (Xi et al., 28 Feb 2026).

3. Annotation pipeline and learning architecture

WristPP relies on a multi-stage annotation pipeline defined in a canonical hand-local coordinate system (Xi et al., 28 Feb 2026). The hand-local frame uses the wrist joint as origin; the $x$-axis runs from wrist to index MCP; the $z$-axis is the palm normal from wrist–index and wrist–pinky vectors; and the $y$-axis completes a right-handed frame. The hand mesh is modeled as

$\mathcal{V}_l = \mathrm{MANO}(\theta,\beta^*)$

with world-space vertices

$$\mathcal{V}_w = R_L^W \mathcal{V}_l + T_L^W = (R_L^{(0)W}\Delta R)\mathcal{V}_l + (T_L^{(0)W}+\Delta T).$$

The optimization variables are

$\Theta=\{\theta,\Delta R,\Delta T,\mathcal{P}_v\}.$

The total annotation loss is written as

$\mathcal{L}_{\text{opti}= \mathcal{L}_{\text{markers}+ \mathcal{L}_{\text{mask}+ \mathcal{L}_{\text{render}+ \mathcal{L}_{\text{anat}.$

Pressure supervision uses differentiable pressure and depth rendering. The pressure loss is

$\mathcal{L}_{\mathrm{press} = \frac{1}{HW}\|\tilde{\mathcal{R}_p - P_{\mathrm{gt}\|_F^2$

and the soft contact probability is

$\hat C=\sigma\!\left(\frac{\varepsilon-d_{\mathrm{rel}{\tau}\right).$

Camera extrinsics are obtained by solving

$\min_{R_{\text{cam}\in SO(3),\, t_{\text{cam}\in\mathbb{R}^3} \sum_{i=1}^N \left\| \pi_{\text{ocam}(R_{\text{cam}X_i+t_{\text{cam})-u_i \right\|_2^2 .$

These components are all explicitly part of the published pipeline (Xi et al., 28 Feb 2026).

The model itself uses a ViT backbone with two sets of 21 learnable tokens, one for pose and one for pressure (Xi et al., 28 Feb 2026). An extrinsics-conditioned branch based on ResNet-50 predicts a 6D rotation representation and translation offset, producing a camera embedding concatenated to both token sets. Pose recovery is formulated as code-index prediction via Hand-VQ-VAE, whose codebook contains 512 embedding vectors in $z$0. For each frame, the encoder outputs 21 latent vectors, each quantized to the nearest codebook vector, and the predicted indices are decoded to recover MANO parameters and the $z$1-vertex hand mesh (Xi et al., 28 Feb 2026).

The pressure branch outputs contact logits and pressure values over the same 778 vertices (Xi et al., 28 Feb 2026). Its loss decomposes into contact classification and foreground-gated pressure regression: $z$2 with

$z$3

and

$z$4

The full training objective is

$z$5

Training uses AdamW, learning rate 1e-5, batch size 32, OneCycle, 6 RTX 4090 GPUs, and about 6 hours total training time (Xi et al., 28 Feb 2026).

4. Reported performance and interaction studies

Offline pose performance is reported as MPJPE = 2.9 mm, MJAE = 3.2°, PA-MPJPE = 2.4 mm, PVE = 3.0 mm, and PA-PVE = 2.6 mm (Xi et al., 28 Feb 2026). For pressure and contact, the paper reports Contact IoU = 0.712, Volumetric IoU = 0.618, Contact Accuracy = 97.1%, and foreground MAE = 10.4 g (Xi et al., 28 Feb 2026). The extrinsics branch achieves rotation error = 2.3°, translation error = 8.9 mm, and 2D reprojection = 13.7 px (Xi et al., 28 Feb 2026). The model runs at about 22 fps on an RTX 4060 GPU (Xi et al., 28 Feb 2026).

The appendix baseline comparison reports a large pose margin over MediaPipe and WiLoR. Overall MPJPE is 46.8 mm for MediaPipe, 15.0 mm for WiLoR, and 3.0 mm for WristP²; overall MJAE is 36.8°, 13.8°, and 3.4°, respectively (Xi et al., 28 Feb 2026). For contact estimation, PressureVision++ is described as conservative and often failing to detect contact, while WristP² wrist-view contact prediction achieves Acc 97.7, Prec 92.6, Rec 98.0, and F1 95.2 (Xi et al., 28 Feb 2026).

Three laboratory studies and one real-world study translate these prediction results into interaction outcomes (Xi et al., 28 Feb 2026). In mid-air Fitts’ law pointing, the virtual air mouse reaches 2.5 bit/s, close to the laptop touchpad’s 2.6 bit/s, though slower than a conventional mouse at 7.5 bit/s. In multi-finger pressure control, the overall success rate = 86.7% with median completion time around 6.1 s. In a virtual pressure-sensitive touchpad task, the system reaches success rate = 98.0%, median total completion time 9.59 s, movement time 3.50 s, and pressure stabilization time 5.52 s. In a 65-inch large-display Whac-A-Mole task, WristP² reports HR 94.47%, error rate 2.07%, reaction time 1.43 s, and score 78.92, outperforming head-mounted RGB camera baselines and producing lower reported fatigue (Xi et al., 28 Feb 2026).

5. Relationship to wrist haptics and wrist perception

WristPP also sits within a broader wrist-perception literature in which sensing or feedback is relocated from the fingertip to the wrist. "Perception of Mechanical Properties via Wrist Haptics: Effects of Feedback Congruence" (Sarac et al., 2022) studies whether fingertip interaction forces should be mapped congruently to wrist skin deformation. The experiment uses 14 participants, Actuonix PQ12-P linear actuators, tracked fingertips at about 200 Hz, CHAI3D updated at 144 Hz, and personalized calibration so normal and shear feedback felt equally intense. The psychometric function is

$z$6

where $z$7 is the point of subjective equality (PSE) and $z$8 is the slope parameter (Sarac et al., 2022).

The principal result is that congruent mapping improved perceptual accuracy, especially by reducing PSE bias for mass and friction, even though JNDs did not significantly improve with congruence (Sarac et al., 2022). For mass discrimination, PSE showed a significant effect of haptic condition, $z$9; for friction discrimination, PSE also showed a significant effect, $y$0 (Sarac et al., 2022). This is relevant to WristPP because it shows that the wrist can encode task-relevant hand-contact information, but the mapping from physical interaction to wrist cue geometry matters.

"Effects of Haptic Feedback on the Wrist during Virtual Manipulation" (Sarac et al., 2022) addresses a related question: whether feedback location and multiplicity on the wrist influence virtual stiffness discrimination. Using a haptic bracelet on the dorsal, ventral, or both sides of the wrist, 12 volunteers compared a 0.3 N/mm reference object against 0.1, 0.2, 0.3, 0.4, or 0.5 N/mm comparison stimuli (Sarac et al., 2022). The main conclusion is that wrist-based feedback can support virtual stiffness perception while leaving the fingertips free, and that there was no significant difference in stiffness perception with stimulation at different and multiple locations overall. This supports the general WristPP premise that useful interaction state can be inferred or conveyed from the wrist even when the fingertip itself is not instrumented (Sarac et al., 2022).

6. Relationship to wrist photoplethysmography and continuous authentication

A separate but methodologically related branch of WristPP-style work uses the wrist for physiological sensing. "Detection of Beat-to-Beat Intervals from Wrist Photoplethysmography in Patients with Sinus Rhythm and Atrial Fibrillation after Surgery" (Tarniceriu et al., 2017) evaluates wrist PPG in 18 patients recovering from surgery, split into 9 with sinus rhythm and 9 with atrial fibrillation. Wrist PPG is recorded with the PulseOn OHR tracker, and ECG reference intervals are derived using Kubios HRV software v2.2 (Tarniceriu et al., 2017). For beat matching, each PPG-detected beat at time $y$1 is matched within

$y$2

where $y$3 is the corresponding IBI (Tarniceriu et al., 2017).

The reported beat-detection performance is 99.44% correct beats, 2.39% extra beats, and 0.56% missing beats for sinus rhythm, versus 97.49%, 2.26%, and 2.51% for atrial fibrillation (Tarniceriu et al., 2017). IBI estimation achieves MAE 7.34 ms in sinus rhythm and 14.31 ms in atrial fibrillation (Tarniceriu et al., 2017). The study concludes that wrist PPG-derived IBI are in close agreement with ECG RRI even in an elderly postoperative population, supporting wrist sensing as a comfortable alternative to chest-based measurement.

The 2025 smartwatch authentication paper extends this wrist-PPG theme from physiology monitoring to biometric identity (Shao et al., 19 Aug 2025). It presents a We-Be Band smartwatch prototype with 4 PPG channels, sampled at 25 Hz, processed in 4-second windows with 50% overlap, and classified using Bi-LSTM + attention (Shao et al., 19 Aug 2025). On the We-Be Dataset of 26 volunteers, the model reaches Average test accuracy: 88.11%, Macro F1: 0.88, FAR: 0.48%, FRR: 11.77%, and EER: 2.76% (Shao et al., 19 Aug 2025). Sensor power is 41.9 mW at 25 Hz, compared with 51.5 mW at 128 Hz and 90.0 mW at 512 Hz, corresponding to 53% lower power than 512 Hz and 19% lower power than 128 Hz (Shao et al., 19 Aug 2025). Although this work is not the camera-based WristPP system, it reinforces the same wearable principle: low-burden wrist sensing can support dense inference continuously in real-world conditions.

7. WristPP in the wider wrist-systems landscape

The significance of WristPP becomes clearer when placed next to wrist-centric control and biomechanics papers. "Continuous Wrist Control on the Hannes Prosthesis: a Vision-based Shared Autonomy Framework" (Vasile et al., 24 Feb 2025) argues that most prosthetic grasp controllers overemphasize fingers and neglect wrist motion, forcing compensatory movements of the elbow, shoulder, and hip. Its eye-in-hand shared-autonomy pipeline divides operation into transport, rotation, and grasping, and uses Image-Based Visual Servoing (IBVS) with the control law

$y$4

plus a partitioned pp-IBVS design to enforce more natural wrist pronation-supination trajectories (Vasile et al., 24 Feb 2025). In simulation, s-IBVS achieved natural configuration success 13/20, while pp-IBVS achieved 20/20, albeit with slower convergence (Vasile et al., 24 Feb 2025). A plausible implication is that WristPP-style wrist instrumentation complements, rather than replaces, intelligent wrist actuation.

Mechanical wrist design papers make a similar point from the hardware side. "Modeling and Control of a Novel Variable Stiffness Three DoFs Wrist" (Milazzo et al., 2023) presents the VS-Wrist, a compact 3-DoF wrist with flexion/extension, radial/ulnar deviation, and pronation/supination, able to triple its stiffness while using only four motors (Milazzo et al., 2023). Its posture reconstruction after calibration reaches an average RMSE of 6.6° with $y$5 (Milazzo et al., 2023). "SoftHand Model-W" (Merritt et al., 1 Apr 2026) integrates a 2-DoF tendon-driven wrist into an anthropomorphic underactuated hand; in an object rotation task, completion time drops from 66 s without wrist actuation to 47 s with wrist actuation, and in cube stacking the wrist-enabled configuration reaches 6/6 stack success and 6/6 reorientation success (Merritt et al., 1 Apr 2026). These results support a consistent systems-level conclusion: wrist capability reduces compensatory motion and expands interaction space.

Biomechanical modeling work extends the same logic to simulation and individualized analysis. "Rapid Development of Efficient Participant-Specific Computational Models of the Wrist" (Andreassen et al., 25 May 2025) presents a workflow using 3DCT and 4DCT, non-linear morphing, GRNN-based cartilage extrusion, and algorithmic ligament generation to create three participant-specific wrist FEMs in under 2 hours per model, with individual simulation runtime of approximately 45 seconds (Andreassen et al., 25 May 2025). This suggests that future WristPP-like systems could be analyzed not only as sensing devices but also as individualized biomechanical interfaces.

Taken together, these papers establish WristPP as part of a larger transition in wrist research: from the wrist as a passive anatomical junction to the wrist as a primary site for dense sensing, redirected haptics, continuous physiological monitoring, shared-autonomy control, and compact mechatronic intelligence (Tarniceriu et al., 2017, Sarac et al., 2022, Sarac et al., 2022, Milazzo et al., 2023, Vasile et al., 24 Feb 2025, Andreassen et al., 25 May 2025, Xi et al., 28 Feb 2026, Merritt et al., 1 Apr 2026).

8. Limitations and open questions

The camera-based WristPP paper explicitly lists several limitations (Xi et al., 28 Feb 2026). The current model is restricted to planar / quasi-planar contact and is “not yet extended to grasping arbitrary objects or curved contact surfaces.” Extreme wrist rotations or bulky objects can cause occlusion. The system is single-hand only and does not yet support bimanual interaction. The prototype remains wearable but “not yet as slim or power-efficient as commercial smartwatches,” and runtime at about 22 fps on RTX 4060 leaves on-device low-power deployment as future work (Xi et al., 28 Feb 2026).

Related literatures identify additional constraints that likely generalize to WristPP-class systems. Wrist haptics papers emphasize that normal and shear intensity require careful calibration and that the wrist is inherently less sensitive than the fingertip (Sarac et al., 2022, Sarac et al., 2022). Wrist PPG papers note the vulnerability of optical inference to motion artifacts and physiological variability, especially in atrial fibrillation or during activity (Tarniceriu et al., 2017, Shao et al., 19 Aug 2025). Prosthetic wrist control papers point out that natural trajectories may impose a convergence-time penalty and that evaluation should eventually include amputees, motion capture, and fatigue measures such as pupil dilation (Vasile et al., 24 Feb 2025).

A plausible implication is that the long-term trajectory of WristPP research will depend less on any single sensing modality than on cross-modal integration. The existing evidence already spans RGB vision, tactile retargeting, optical pulse sensing, variable-stiffness mechanics, and participant-specific biomechanics. What unifies these efforts is a consistent design proposition: the wrist can support high-value inference and control without monopolizing the fingertips, external workspace, or upper-limb kinematics (Tarniceriu et al., 2017, Sarac et al., 2022, Sarac et al., 2022, Milazzo et al., 2023, Vasile et al., 24 Feb 2025, Andreassen et al., 25 May 2025, Xi et al., 28 Feb 2026).