EgoPhys: First-Person Physiology & World Modeling

• EgoPhys is an umbrella term for first-person systems that infer latent internal and physical states using egocentric sensors and structured modeling techniques.
• It integrates advanced techniques such as eye-tracking-based heart rate estimation, affect recognition, and kinematics-based world modeling for motion forecasting and controllable synthesis.
• EgoPhys demonstrates practical applications in digital twin generation and human–object interaction simulation, impacting both research and real-world robotics.

EgoPhys denotes an emerging egocentric research agenda that couples first-person sensing with internal physiology or physically grounded world modeling. In the cited literature, the term has two closely related uses. In the physiological-sensing line, EgoPhys refers to egocentric physiological sensing for modeling affect, personality, and behavior, and includes heart-rate estimation from eye-tracking cameras and multimodal affect recognition from smartglasses (Braun et al., 28 Feb 2025, Jammot et al., 25 Oct 2025). In the world-modeling line, it refers more conceptually to egocentric physical/world modeling grounded in kinematics, action-conditioned prediction, controllable first-person synthesis, and deformable-object simulation; the framework titled "EgoPhys" makes this interpretation explicit by learning deformable physical digital twins from egocentric RGB-only video (Patel et al., 2 Aug 2025, Tran et al., 14 May 2026, Li et al., 13 Mar 2026, Gu et al., 25 May 2026, Kim et al., 15 Jun 2026).

1. Scope and operationalizations

The recent literature does not use EgoPhys as a single narrowly defined task. Instead, it spans egocentric physiological sensing, action-conditioned world modeling, physically grounded video synthesis, and deformable simulation. This suggests that EgoPhys is best understood as an umbrella term for first-person systems that attempt to infer or simulate latent internal or physical state, rather than merely recognize visible activity.

Work	Operationalization	Core signals
egoPPG (Braun et al., 28 Feb 2025)	heart rate estimation from eye-tracking cameras	inward-facing eye-tracking video, contact PPG, ECG
egoEMOTION (Jammot et al., 25 Oct 2025)	affect and personality recognition	eye-tracking video, POV camera, head IMU, PPG, ECG, EDA, respiration
UniEgoMotion (Patel et al., 2 Aug 2025)	motion reconstruction, forecasting, and generation	egocentric images, device SLAM trajectory
EgoExo-WM (Tran et al., 14 May 2026)	whole-body action-conditioned egocentric world model	egocentric video latents, SMPL pose increments, converted exocentric data
EgoHOI (Li et al., 13 Mar 2026)	egocentric HOI world model	first-frame RGB, hand kinematics, metric head motion
E$^3$C (Gu et al., 25 May 2026)	controllable egocentric video generation	context frames, camera path, ego–exo pose control, 3D memory
EgoPhys (Kim et al., 15 Jun 2026)	deformable physical digital twin generation	single egocentric RGB video, masks, tracks, 4D point cloud

Across these works, the common design pattern is not raw video prediction alone. Instead, the systems inject structure: physiological proxies, whole-body pose increments, head-centric canonicalization, hand kinematic renders, 3D environmental memory, or state-conditioned spring stiffness fields. The principal technical distinction among them is which latent state they seek to recover: heart-rate and affective state, articulated body motion, first-person visual futures, or deformable-object mechanics.

2. Physiological state estimation and affect modeling

In the physiological line of EgoPhys, the core claim is that egocentric systems should detect the wearer’s internal state because physiological state influences cognitive performance, attention, situational responses, emotions, stress, fatigue, and alertness. The paper "egoPPG" defines a novel task for egocentric vision systems: extracting heart rate from built-in sensors, specifically eye-tracking cameras, without requiring additional or dedicated hardware (Braun et al., 28 Feb 2025).

The egoPPG system uses inward-facing eye-tracking video frames captured by Project Aria, monochrome with IR illumination, and estimates a photoplethysmogram from regions around the eyes. The physical model is stated as

$$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$

with diffuse reflection decomposition

$$\boldsymbol{v_d}(t)=\boldsymbol{u_d} d_0+\boldsymbol{u_p} p(t),$$

where $p(t)$ is the desired BVP signal. EgoPulseFormer preprocesses 30 fps eye-tracking video in temporal windows of $T=128$ frames, standardizes each frame, and uses standardized consecutive frame differences as network input. Its backbone is a 3D CNN based on PhysNet with spatial attention modules that focus on stable, high-SNR skin regions around the eyes and down-weight low-SNR eye regions subject to frequent motion and blinks. On egoPPG-DB, which contains 13+ hours of synchronized eye-tracking video, contact PPG, and ECG from 25 participants performing office work, kitchen work, dancing, cycling, and walking, the model reports MAE $=8.82$ bpm, RMSE $=12.03$, MAPE $=10.82$, and $r=0.81$, with HR spanning $44$–$$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$0 bpm (Braun et al., 28 Feb 2025).

The significance of egoPPG is not limited to physiological measurement. The paper reports that augmenting downstream EgoExo4D proficiency estimation with tracked HR values improves top-1 accuracy from $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$1 to $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$2 for Ego only, a $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$3 relative improvement, and from $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$4 to $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$5 for Ego + Exo, an $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$6 improvement. This positions heart rate as contextual state information that can improve egocentric video understanding rather than as an isolated biosignal (Braun et al., 28 Feb 2025).

The dataset "egoEMOTION" broadens this physiological interpretation of EgoPhys from heart rate to affect and personality. It couples first-person visual signals from smartglasses with concurrent physiological sensing and dense self-reports of affect and emotion plus trait-level personality. The dataset contains over 50 hours from 43 participants, using Meta Project Aria with eye-tracking videos at $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$7 per eye and 90 fps, POV RGB at $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$8 and 10 fps, head IMUs at 1000 Hz and 800 Hz, nosepad PPG at 128 Hz, Shimmer3 PPG and EDA at 256 Hz, ECG at 1024 Hz, respiration at 400 Hz, and an external face webcam at $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$9 and 60 fps (Jammot et al., 25 Oct 2025).

egoEMOTION defines three benchmark tasks: continuous affect classification for valence, arousal, and dominance; discrete emotion classification; and trait-level personality inference. Its central empirical result is that head-mounted egocentric signals can match or surpass traditional physiology-only baselines in real-world affect prediction. For continuous affect, mean F1 is $$\boldsymbol{v_d}(t)=\boldsymbol{u_d} d_0+\boldsymbol{u_p} p(t),$$0 for all modalities fused, $$\boldsymbol{v_d}(t)=\boldsymbol{u_d} d_0+\boldsymbol{u_p} p(t),$$1 for egocentric glasses only, and $$\boldsymbol{v_d}(t)=\boldsymbol{u_d} d_0+\boldsymbol{u_p} p(t),$$2 for wearable physiological devices only. For discrete emotions, egocentric glasses only and all modalities fused both achieve mean F1 $$\boldsymbol{v_d}(t)=\boldsymbol{u_d} d_0+\boldsymbol{u_p} p(t),$$3, compared with $$\boldsymbol{v_d}(t)=\boldsymbol{u_d} d_0+\boldsymbol{u_p} p(t),$$4 for fused wearables. For personality, all modalities fused achieve mean F1 $$\boldsymbol{v_d}(t)=\boldsymbol{u_d} d_0+\boldsymbol{u_p} p(t),$$5, compared with $$\boldsymbol{v_d}(t)=\boldsymbol{u_d} d_0+\boldsymbol{u_p} p(t),$$6 for egocentric glasses only and $$\boldsymbol{v_d}(t)=\boldsymbol{u_d} d_0+\boldsymbol{u_p} p(t),$$7 for wearables only. Classical methods also outperform the deep learning baselines across all three tasks (Jammot et al., 25 Oct 2025).

A common misconception is that EgoPhys in this sense is equivalent to conventional physiological monitoring. The empirical record here is narrower and more specific: eye-tracking video, pupil diameter, gaze, blink dynamics, per-frame eye intensity, Fisherface projections, LBP-TOP micro-expressions, and head IMU statistics can be more informative than peripheral physiology alone for some affective tasks in real-world egocentric settings (Jammot et al., 25 Oct 2025).

3. Kinematics-grounded motion and world models

A second major interpretation of EgoPhys centers on egocentric world modeling with structured, interpretable action spaces. "UniEgoMotion" formulates egocentric motion reconstruction, forecasting, and generation from first-person images, optionally with the head-mounted device’s SLAM trajectory, without relying on explicit 3D scene input (Patel et al., 2 Aug 2025). Its core technical move is a head-centric motion representation tailored to egocentric devices. If the head pose in the world is $$\boldsymbol{v_d}(t)=\boldsymbol{u_d} d_0+\boldsymbol{u_p} p(t),$$8, a world-space joint position is mapped into the head-centric frame by

$$\boldsymbol{v_d}(t)=\boldsymbol{u_d} d_0+\boldsymbol{u_p} p(t),$$9

The canonical head transform removes pitch, roll, and height and keeps yaw, while the residual trajectory is represented as $p(t)$0. The reported effect is lower foot sliding and floor penetration than pelvis-centric or chain-preserving canonicalizations (Patel et al., 2 Aug 2025).

UniEgoMotion trains a unified conditional diffusion model on EE4D-Motion, which yields 208 hours of paired egocentric video and pseudo-ground-truth 3D motion, with 110+ hours remaining for training after quality filtering. On reconstruction, it reports MPJPE $p(t)$1 m versus $p(t)$2 m for AvatarPoser, Foot Sliding $p(t)$3 versus $p(t)$4, Foot Contact $p(t)$5 m versus $p(t)$6, Semantic Sim $p(t)$7 versus $p(t)$8, and FID $p(t)$9 versus $T=128$0. On forecasting after 2 seconds of observation, it reports MPJPE $T=128$1 m and FS $T=128$2. On generation from a single image, it reports MPJPE $T=128$3 m and FS $T=128$4, with the paper emphasizing realism and reduced sliding rather than minimum MPJPE alone (Patel et al., 2 Aug 2025).

"EgoExo-WM" extends this line from motion generation to planning-oriented egocentric world models. It defines the action $T=128$5 as a structured whole-body control signal derived from SMPL body pose, with root translation and per-joint relative rotations:

$T=128$6

The world model predicts the next egocentric visual latent as

$T=128$7

with $T=128$8 extracted by DINOv3-L and prediction performed by a CDiT-L/2 video transformer. A wrist-consistency head predicts next-frame wrist heatmaps, and the training loss is

$T=128$9

The distinctive contribution is not only the action space but also the conversion of in-the-wild exocentric video into egocentric training data through EgoX-Body, which uses a head-anchored camera constructed from eye joints and a forward offset of approximately $=8.82$0 m (Tran et al., 14 May 2026).

EgoExo-WM mixes converted exocentric data with Nymeria and reports consistent improvements in both prediction and planning on HOMAGE, LEMMA, and Ego-Exo4D subsets. On HOMAGE, EgoExo-WM achieves $=8.82$1 final 2s latent L2, $=8.82$2 average L2, and wrist PCK@20 $=8.82$3, outperforming Ego-WM and naive EgoExo-WM. In planning, HOMAGE improves from UniEgoMotion $=8.82$4 MPJPE/Wrist MPJPE to $=8.82$5 with EgoExo-WM; similar gains are reported on LEMMA, Bike, and Cooking. The paper explicitly notes that dynamics and contact are not explicitly modeled and that physical plausibility during planning is induced by sampling candidate motions from UniEgoMotion rather than by explicit $=8.82$6 terms (Tran et al., 14 May 2026).

A misconception in this area is that any egocentric video predictor is a world model in the physical sense. The cited works distinguish action-conditioned, kinematically grounded prediction from generic conditional generation: explicit pose increments, forward kinematics, and head-anchored camera models are treated as essential for planning and embodied control (Patel et al., 2 Aug 2025, Tran et al., 14 May 2026).

4. Photorealistic interaction synthesis and controllable egocentric generation

"EgoHOI" addresses egocentric Human–Object Interaction world modeling as an action-driven simulator that predicts physically grounded first-person rollouts of hand–object interactions. The state transition is

$=8.82$7

where $=8.82$8 is hand kinematics from reconstructed MANO-based hand meshes rendered in the egocentric camera, and $=8.82$9 is metric head motion from calibrated camera intrinsics and extrinsics. The model uses a latent video VAE and a Wan-DiT backbone, together with three physics-informed embedding adapters: Hand Kinematic Embeddings, Ego-Motion Embeddings, and Object-Entity Embeddings. The paper is explicit that EgoHOI does not directly output contact graphs or explicit force/physics variables, and that contact realism is encouraged implicitly by the embedding design rather than by explicit contact or collision losses (Li et al., 13 Mar 2026).

On HOT3D, EgoHOI reports strong gains over Wan, Cosmos 2B, Cosmos 14B, and Uni3C. For frame prediction it reports PSNR $=12.03$0, SSIM $=12.03$1, LPIPS $=12.03$2, and Object-CLIP $=12.03$3. For ego-motion consistency it reports ATE $=12.03$4, RRE $=12.03$5, and RPE $=12.03$6. For kinematic fidelity it reports Missing Ratio $=12.03$7, MPJPE $=12.03$8, and RMSE $=12.03$9. The ablations show that HKE primarily improves kinematic fidelity, EME improves ego-motion consistency, and OEE stabilizes object identity and appearance, reducing OPE to $=10.82$0 and OOE to $=10.82$1 in the full model (Li et al., 13 Mar 2026).

E$=10.82$2C treats EgoPhys from a complementary angle: controllable egocentric video generation grounded by a persistent 3D environmental memory and joint ego–exo human motion controls. It constructs a semi-dense point cloud memory

$=10.82$3

where each point stores world-space position, pooled RGB, and a per-point appearance descriptor from video-VAE features. Rendering this memory into target viewpoints produces view-aligned conditioning. Human dynamics are separated: exo people are controlled by 2D skeleton renderings, while the camera wearer is specified by 3D body joints, 6DoF wrist poses, and persistent ego-motion tokens for cross-attention (Gu et al., 25 May 2026).

On Nymeria test, E$=10.82$4C reports FVD $=10.82$5, LPIPS $=10.82$6, PSNR $=10.82$7, SSIM $=10.82$8, TErr $=10.82$9 cm, RErr $r=0.81$0, Obj-F1 $r=0.81$1, Obj-mIoU $r=0.81$2, Exo-F1 $r=0.81$3, PCK@10\% $r=0.81$4, Hand-F1 $r=0.81$5, and Hand-mIoU $r=0.81$6. The ablations indicate that point-cloud rendering produces the largest gain in fidelity and camera control, while per-point appearance features, exo skeletons, drawn ego pose, and encoded ego tokens progressively improve texture, exo adherence, and ego control (Gu et al., 25 May 2026).

These papers sharpen an important controversy around the word “physical.” In EgoHOI, physical fidelity is induced by metric and kinematic priors but not by explicit contact, force, friction, or penetration losses. In E$r=0.81$7C, physical grounding comes from geometric consistency via camera projection and 3D memory, with no auxiliary camera or object losses. The term therefore does not imply a full physics engine in every case; it often denotes stronger structural constraints on first-person generation (Li et al., 13 Mar 2026, Gu et al., 25 May 2026).

5. Deformable-object digital twins

The framework titled "EgoPhys" specializes the broader agenda to deformable-object mechanics learned from egocentric RGB-only video. Its target is a controllable deformable digital twin constructed from a single 7 s egocentric RGB video captured by Meta Project Aria at 30 fps and $r=0.81$8, without multi-view RGB-D capture or per-spring test-time optimization (Kim et al., 15 Jun 2026).

The pipeline first converts video into an egocentric 4D point cloud. Grounded-SAM2 initializes and propagates object and hand/manipulator masks; CoTracker3 computes dense 2D tracks; VGGT predicts per-pixel world coordinates $r=0.81$9 and confidence $44$0; tracked pixels $44$1 are lifted independently per frame, keeping only confident, in-range points according to

$44$2

with typical thresholds $44$3 and depth range $44$4–$44$5 m. After filtering, control points are sampled via farthest-point sampling, approximately 30 controller points, and geometry can be optionally completed with TRELLIS when occlusion leaves large gaps (Kim et al., 15 Jun 2026).

The deformable object is modeled as a spring-mass graph with structural, bending, and controller springs as appropriate. Dynamics are integrated with explicit Euler:

$44$6

CMA-ES, approximately 50 generations, estimates graph topology and coarse global physical parameters by minimizing geometry and motion error over rollouts. The central innovation is then a codebook-based physics prior that predicts dense, state-conditioned per-spring stiffness fields without test-time refinement. Each spring uses a log-stiffness offset relative to the coarse global stiffness:

$44$7

The offset is produced by a sign-aware dynamic codebook with separate tension and compression banks:

$44$8

The default configuration uses $44$9 prototypes, temperature $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$00, an MLP of width 64, and Adam with learning rate $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$01 (Kim et al., 15 Jun 2026).

Empirically, EgoPhys reports strong gains over adapted PhysTwin and Spring-Gaus baselines. In reconstruction and resimulation, it achieves CD $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$02, TE $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$03, IoU $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$04, PSNR $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$05, SSIM $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$06, and LPIPS $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$07. In future prediction, it reports CD $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$08, TE $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$09, IoU $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$10, PSNR $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$11, SSIM $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$12, and LPIPS $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$13. In zero-shot generalization to unseen objects and interactions, EgoPhys reports CD $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$14, TE $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$15, IoU $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$16, PSNR $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$17, SSIM $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$18, and LPIPS $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$19, outperforming PhysTwin while avoiding any test-time spring optimization. The codebook also outperforms dense per-sequence refinement while running in approximately $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$20 min/sequence; dense refinements with 25–200 steps are reported as $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$21–$$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$22 slower and consistently worse (Kim et al., 15 Jun 2026).

The framework is further deployed on a real xArm6 robot in lifting and pulling tasks. From a single egocentric human video, EgoPhys builds the twin, plans with MPPI, transfers waypoints via a calibrated sim-to-robot transform, and executes with a simple gripper. Reported object-configuration error decreases are $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$23 for fox lift, $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$24 for green monster lift, and $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$25 for Doraemon pull (Kim et al., 15 Jun 2026).

6. Evaluation regimes, limitations, and future directions

The evaluation culture around EgoPhys is heterogeneous because the underlying latent state differs by subfield. Physiological EgoPhys uses MAE, RMSE, MAPE, Pearson $$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$26, and F1 scores; motion models use MPJPE, foot sliding, foot contact, Head Rotation Error, Semantic Similarity, and FID; egocentric video models use FVD, LPIPS, PSNR, SSIM, wrist PCK, ATE, RRE, RPE, Object-CLIP, OPE, and OOE; deformable simulation uses CD, track error, IoU, and rendering metrics (Braun et al., 28 Feb 2025, Jammot et al., 25 Oct 2025, Patel et al., 2 Aug 2025, Li et al., 13 Mar 2026, Gu et al., 25 May 2026, Kim et al., 15 Jun 2026). This diversity reflects different targets rather than a lack of methodological coherence.

Several limitations recur. egoPPG reports highest errors during high motion and elevated HR activities such as dancing and cycling, and performance degrades at 10 fps unless interpolated back to 30 fps (Braun et al., 28 Feb 2025). egoEMOTION notes retrospective per-task labels, limited representation of low-arousal, low-valence states, young-adult demographic skew, and the risk that IMU features may confound affect with task-related motor activity (Jammot et al., 25 Oct 2025). UniEgoMotion depends on pseudo-ground-truth annotations and a floor-aligned canonical frame, which becomes brittle on non-flat terrain or under severe occlusion and blur (Patel et al., 2 Aug 2025). EgoExo-WM does not explicitly model dynamics or contact and omits hand articulation from the action space (Tran et al., 14 May 2026). EgoHOI does not explicitly optimize contacts, forces, friction, or penetration, and current metrics remain proxies for true physical plausibility (Li et al., 13 Mar 2026). E$$\boldsymbol{C(t)}=I(t) (\boldsymbol{v_s}(t)+\boldsymbol{v_d}(t))+\boldsymbol{v_n}(t),$$27C does not explicitly model non-human scene dynamics, and sparse memory in textureless areas can underconstrain appearance (Gu et al., 25 May 2026). EgoPhys itself uses simplified contact and friction, enables self-collision only for cloth-like objects, and presents robot experiments as proof-of-concept rather than broad closed-loop evaluation (Kim et al., 15 Jun 2026).

Ethics and privacy are especially salient in the physiological and egocentric-sensing branches. egoPPG reports participant consent and ethics approval, but additional privacy analyses or data-protection measures beyond consent and approval are not specified (Braun et al., 28 Feb 2025). egoEMOTION is more explicit: it cites ETH Zürich Ethics Commission no. 23 ETHICS-008, notes that egocentric, eye-tracking, and external videos are inherently identifiable, and restricts dataset access to permanent academic staff via a Data Transfer and Use Agreement (Jammot et al., 25 Oct 2025).

The near-term trajectory of EgoPhys is correspondingly clear in the cited work. The physiological branch points toward richer multimodal internal-state modeling, including HR, affect, emotion, and personality, with better robustness under motion and broader demographic coverage (Braun et al., 28 Feb 2025, Jammot et al., 25 Oct 2025). The kinematic and world-modeling branch points toward richer action spaces, stronger contact and dynamics modeling, and planning with more physically plausible proposals (Patel et al., 2 Aug 2025, Tran et al., 14 May 2026). The generative branch points toward causal first-person simulators that remain controllable under severe occlusion and rapid ego-motion (Li et al., 13 Mar 2026, Gu et al., 25 May 2026). The deformable-object branch points toward larger egocentric datasets, improved contact and friction modeling, and closed-loop sim-to-real control from videos people already capture (Kim et al., 15 Jun 2026).

Taken together, these works define EgoPhys as a shift in egocentric research from appearance-centric recognition toward latent-state estimation and simulation. Whether the latent state is heart rate, affective disposition, articulated body motion, hand–object interaction dynamics, scene-consistent future video, or deformable-object mechanics, the unifying premise is that first-person systems become more useful when they model what is physically or physiologically driving the observation, rather than only what is visible in the observation itself.