VOIR · RESEARCH REGISTRY · OPEN CORPUS
Eighty-eight short papers on the techniques and instruments shaping augmented reality, immersive computing, and on-device perception. The lab observes; the lab summarizes; the lab declines to disclose how the methods are combined behind the wall.
88 of 88 entries shown
VLMs generate fluent text whether or not the underlying perception supports it. The paper examines refusal training and uncertainty-aware decoding for spatial tasks. The lab argues that interactive AR favors *short refusals* over long confabulations, and proposes a per-token entropy gate.
First-launch re-localization is the most user-hostile moment in persistent AR: nothing is known about the room until the device sweeps it. The paper digests four cold-start strategies: dense priming sweeps, gravity-aligned mini-maps, server-side coarse priors, and IMU-only blind tracking until the first feature lock. None is satisfying alone; the lab argues for a hybrid with explicit user signaling on the first session.
Reinforcement learning is tempting for spatial tasks but is rarely the cheapest answer. The paper catalogues five tasks where the lab considered RL and chose classical control or imitation instead, with rationale. A decision tree is offered: if the reward is dense, the demonstrations exist, and the simulator is faithful, RL is reasonable; otherwise, it is not.
Eye tracking on phone front cameras is feasible at coarse resolution; gaze prediction within a held device is dominated by head pose, not eyeball pose. The paper reviews mobile-gaze methods (GazeCapture, OpenGaze) and their characteristic failure modes.
Knowledge distillation has become a default tool for shipping research-grade models to mobile. The paper reviews logit distillation, feature distillation, attention transfer, and self-distillation. A finding: distillation works best when the student architecture is similar to the teacher; cross-family distillation is brittle.
Volumetric AR effects (smoke, light shafts, transparency) remain expensive on mobile. The paper indexes density-grid rendering, sparse voxel approaches (NSVF), and signed-distance volumes. Findings: small volumes are cheap; large volumes are not yet tractable.
Polarization-aware imaging is mature in microscopy, fragile in mobile. The paper reviews polarimetric phones (Sony IMX-class with pixel polarizers), specular separation (PolDepth), and shape-from-polarization. Filed as a watchlist item: not deployed, of interest.
Real-time on-device 3DGS rendering crossed the 30 fps threshold on flagship-class mobile NPUs in late 2025. The paper sketches the bottlenecks that remain — splat sort, alpha compositing, memory bandwidth on tiled GPUs — and the techniques that have been tried (hierarchical splat LODs, view-frustum culling, adaptive radius). It does not describe shipped lab pipelines.
Sub-2B parameter LLMs (Phi-3.5-mini, Gemma-2B, Llama-3.2-1B, Qwen-2.5-1.5B) now fit on flagship phones with 1–3 GB of RAM. The paper reports tokens-per-second under thermal load, INT4 quantization, and KV-cache compression. The privacy benefit is substantial; the latency cost is not yet zero.
Indoor magnetic fields are highly non-uniform and surprisingly stable — a fingerprint that is reproducible to ~30 cm in steel-framed buildings. The paper reviews the magnetic-fingerprint literature (IndoorAtlas-style approaches, learned magnetic descriptors) and reports a fusion experiment combining VIO with a continuous magnetic prior. The fusion adds 2 ms of compute per frame and reduces long-session drift in featureless corridors by ~38%.
VLMs are increasingly used as teachers for downstream task-specific models. The paper reviews CLIP→detector distillation (RegionCLIP, GLIP), VLM→segmentation (LISA), and VLM→tracker pipelines. A note: cross-modal distillation is more sensitive to the prompt distribution than to the teacher's underlying quality.
On-device LLM inference at >10 tok/s on flagship phones is now routine for sub-2B models; latency at first-token remains the dominant user-perceived delay. The paper reviews KV-cache compression, speculative decoding, and prompt prefix caching. Findings: prefix caching is the highest-leverage trick for interactive AR.
Success rate is a thin metric for spatial agents. The paper proposes auxiliary metrics — path efficiency, exploration coverage, obstacle margin, time-to-first-error — and reports them on three open benchmarks. A digest of how published work would have changed under richer evaluation is included.
3D Gaussian Splatting renders beautifully but produces clouds, not surfaces. The paper reviews mesh-extraction strategies (SuGaR, 2DGS, GS2Mesh, Poisson reconstruction over splat centroids). Each pays a different fidelity tax. The lab notes the open problem of producing watertight, edited, semantically labeled meshes from a splat cloud at interactive rates.
Voice is the connective tissue of spatial UI when both hands are committed. The paper reviews on-device wake-word detection, intent recognition, and the design tension between push-to-talk and always-listening modes. Findings: ambient voice mode is rarely justified.
Mobile hand-pose models (MediaPipe Hands, RTMPose-Hand, HandFormer) achieve sub-pixel accuracy on isolated hands. Two-hand interaction collapses accuracy by 30–50% under self-occlusion. The paper reviews occlusion-aware training data (InterHand2.6M, AssemblyHands) and inference-time decoupling.
When two devices hold anchors in the same physical room, their anchor poses can be aligned by visual co-observation, ICP on partial reconstructions, or by exchanging encrypted descriptor sketches. The paper enumerates threat models for each. The lab finds that observation-only alignment (no descriptor exchange) is feasible to ~3 cm registration on shared planar surfaces. Privacy considerations are discussed in companion paper RP-061.
Modern phones host 2–4 cameras with different intrinsics. The paper reviews multi-camera factory calibration, ad-hoc stereo from wide/tele pairs, and learned cross-camera depth (DUSt3R, MASt3R). Findings: learned methods are now competitive with classical stereo on close-range subjects.
Beyond text and image, multimodal agents are beginning to absorb gesture as a query channel. The paper reviews tap-on-image, point-on-image, and free-hand input as VLM prompts. Pointing tokens (Molmo-style) are evaluated as a substrate for spatial gesture.
Fixed-foveated rendering is shipping on mobile-class HMDs; eye-tracked foveation requires gaze and is rarer. The paper reports compute savings on representative scenes and notes the strong dependency on the foveation kernel design. Aggressive foveation creates visible boundaries during head motion.
Running a world-model rollout at interactive rates on a phone NPU is feasible only with aggressive latent compression. The paper sketches the trade-offs between rollout horizon, latent dimensionality, and rollout count under a 16 ms budget. The lab does not deploy world-model rollouts in production; the survey is for completeness.
Human pose estimation on mobile splits into top-down (HRNet-class) and bottom-up (OpenPose-class) families. The paper benchmarks on a 30-subject home-fitness dataset and finds bottom-up wins under occlusion; top-down wins on isolated subjects. Latency budgets force a hybrid in practice.
Mobile ML compiler stacks have consolidated: Core ML (Apple), QNN (Qualcomm), TensorFlow Lite + XNNPACK, ONNX Runtime, and PyTorch ExecuTorch. The paper benchmarks the same model across each and reports the effort cost of supporting all paths. Findings: ONNX Runtime is the most predictable; vendor-specific paths are the fastest.
Modern SLAM frontends assume locally planar motion, which collapses on stairwells. The paper reviews three corrections — explicit floor estimation, IMU-only altitude during stair regions, and barometric anchoring — across a 14-building dataset. A simple fusion of barometric altitude with visual scale recovery yields the best vertical accuracy at the cost of barometer warm-up.
Recent work on privacy-preserving structure-from-motion (DP-SfM, line-cloud SfM, hashed-feature SfM) attempts to map a space without retaining recoverable imagery. The paper reviews where these methods succeed (sparse texture, repetitive environments) and where they collapse (fine detail, dynamic objects).
By early 2026 the architectural debate between implicit (NeRF-style) and explicit (point-, splat-, voxel-based) representations has cooled. Explicit representations win on speed, editability, and storage; implicit representations win on continuous derivatives and surface extraction. The paper argues neither is a winner; the production answer is a translation step between the two on demand.
AR comfort literature is large and contradictory. The paper synthesizes a short index: vection cues, vergence-accommodation conflict, FoV transitions, and display-latency variance. Each is mapped to a measurable display or render parameter.
VLM hallucinations are predictable on indoor scenes: confident misnaming of generic furniture, fabricated text on labels, and hallucinated counts. The paper proposes a hallucination-detection protocol using consistency over multiple crops and entropy thresholds.
Long-horizon embodied tasks (clean the room, prepare the workspace) defeat flat policies. The paper reviews hierarchical RL (FuN, options framework, HiTUT) and language-as-policy approaches (SayCan, Code-as-Policies, VoxPoser). The temptation to use an LLM as the high-level controller is acknowledged; failure modes are catalogued.
Smartphone SoCs sustain peak compute for 30–90 seconds before thermal throttling forces a step-down. The paper measures the throttle curve under a representative AR workload (camera + detection + render at 30 fps) and reports the steady-state compute envelope. Steady-state is roughly 60% of peak.
Most reconstruction pipelines assume the scene is static. People in the frame appear as ghost geometry. The paper inventories solutions: explicit human masks (Mask-RCNN, SAM-class), motion-segmentation, neural decomposition (D-NeRF, K-Planes-Dynamic). Notes a finding that two-second masking pre-pass costs less compute than letting the optimizer fight ghost residuals.
Phone IMUs drift with chassis temperature. The paper reports thermal-drift profiles from a 60-minute heating session and proposes online thermal-bias estimation. Worth filing: drift is non-monotonic across SoC throttle events.
Open-vocabulary detection accepts a free-text query and returns boxes. Owl-ViT, GroundingDINO, YOLO-World, and DETR-class variants converge on similar capabilities at very different cost points. The paper indexes per-class accuracy on furniture, signage, and tools. The query-language design is found to dominate downstream behavior.
Render-on-warp (ATW), late-latch poses, and asynchronous reprojection are the three rendering tricks that hide latency on AR displays. The paper reviews each and reports the residual visible artifacts at 90 Hz vs. 120 Hz refresh. Higher refresh narrows the artifact window non-linearly.
A persistent anchor is, in principle, a sparse pose; in practice the descriptor cloud surrounding it can be inverted to recover textured geometry. The paper reviews descriptor inversion attacks (TURF, MapAttack, Pittinverse) and counter-measures (descriptor truncation, randomized features, encrypted anchors).
Below ~10 lux, descriptor extractors based on FAST/BRIEF degrade rapidly; learned descriptors (SuperPoint, R2D2, DISK) hold longer but consume an order of magnitude more inference. The lab notes a counterintuitive observation: increasing exposure helps the front-end but hurts the back-end (motion blur dominates re-projection error). A short discussion of event-camera complements is included.
Video-VLMs (Video-LLaVA, VideoChat2, LongVA, LLaVA-OneVision) extend single-frame reasoning across temporal context. The paper benchmarks them on indoor activity-recognition and event-localization tasks. Performance drops sharply past 30-second windows; positional-encoding tricks help only modestly.
Motion-to-photon latency below ~20 ms is the threshold for comfortable AR. The paper decomposes the budget across capture, detection, pose update, render, and display. The current bottleneck on mobile-class compute is the detection-and-pose stage; the display is rarely the limiting factor.
INT8 post-training quantization (PTQ) is a solved problem for most CNN-class detectors; INT4 is not. The paper reviews QAT strategies (LSQ, OmniQuant, GPTQ-class) and reports calibration-set sensitivity. A finding worth filing: detector heads and backbone sometimes prefer different bit widths.
Behavior cloning (BC) remains the simplest imitation-learning baseline; conservative offline-RL methods (CQL, IQL, AWAC) close the gap to online RL when demonstrations are scarce. The paper reports per-task efficiency on a 200-demo handheld interior dataset. Demo quality dominates demo quantity.
Recovering a coarse scene layout — walls, floor, ceiling, openings — from a single RGB frame is a strongly assumed problem. Methods that assume Manhattan-world geometry (LayoutNet, HorizonNet, AtlantaNet) fail on non-orthogonal architectures common in older buildings. The paper notes a lab finding: the failure manifold of these methods correlates with construction era more than it correlates with image quality.
Time-of-flight (ToF) and structured-light sensors on mobile (iPad Pro / iPhone Pro LiDAR-class, Android ToF) provide low-resolution but absolute-scale depth. The paper reviews their use in VIO scale recovery, mesh seeding, and people-segmentation. Conclusion: helpful, not transformative.
When the device returns to a previously visited region without a place-recognition prior, the residual error per loop accumulates against the inverse covariance of the back-end pose graph. Sparse bundle adjustment (SBA) and incremental smoothing (iSAM2) handle this in classical pipelines, but the paper notes both struggle with degenerate motion (pure rotation) common in headset use. A factor-graph perspective with on-manifold updates is sketched.
Mirror surfaces in AR scenes are routinely mishandled. The paper reviews reflection-probe methods, screen-space reflections (SSR), and ray-marched reflections at mobile rates. Findings: the gap between physically correct and acceptable is wide; users tolerate qualitative reflections.
AR sessions frequently keep the microphone open. The paper reviews on-device wake-word detection (Porcupine-class, RNN-T-class), VAD with explicit user signaling, and audio hashing for command recognition without transcript retention. The trade-off is intent latency vs. retained audio.
Object detectors trained on COCO underperform on interiors by 12–18 mAP points. The paper reviews fine-tuning, domain randomization, and synthetic data (SUN-RGBD, Hypersim, ScanNet) as remediation. Synthetic data narrows but does not close the gap; the missing factor is texture diversity, not geometry.
Reward shaping in 3D navigation is a perennial source of policy pathology. The paper catalogs shaping bugs that have appeared in published work (loop-back exploits, time-pressure compensation, distance-only rewards rewarding wall-hugging). A defensive checklist is offered.
Hitting 30 fps on phone-class compute is an exercise in subtraction. The paper digests the typical frame: capture, ISP, detection, segmentation, pose, render, composite, display. Each stage has a budget; the paper reports the typical envelope on flagship 2025 devices.
Photogrammetric reconstruction from handheld phone footage fails in three ways: specular surfaces, transparent surfaces, and uniform-textured walls. The paper reviews mitigations (multi-view photometric stereo, polarimetric capture, learned depth priors) and notes that none is a complete answer. A taxonomy of interior failure cases by surface type is provided as a lookup.
Waveguide optics, geometric optics, and free-form combiners coexist in 2025-era HMDs. The paper indexes field-of-view, eye-relief, eyebox, and ambient-light tolerance. Worth filing: no current display class clears 60° FoV at sustained day-bright contrast.
VLMs have been wired into tool-use pipelines (function calling, structured output, JSON-mode). The paper reviews indoor tasks that benefit (calibration, measurement, captioning a known scene) and tasks that do not (real-time tracking). A short note on output-token budgeting under interactive constraints is included.
Real-time light estimation in AR is dominated by spherical harmonic regression from the live frame. The paper compares ARKit's environment textures, PointAR-style estimators, and learned HDR-from-LDR methods. A note: harsh point lights remain difficult.
Real-time face blurring on AR streams depends on the underlying face detector. The paper reviews YuNet, BlazeFace, RetinaFace-Mobile, and Mediapipe Face Detector under occlusion, profile views, and low light. The privacy guarantee is bounded by the detector's recall, not its precision.
The magnetometer is the most under-used sensor on a smartphone. Indoor magnetic anomalies are stable and globally unique. The paper reviews magnetic-fingerprint navigation (IndoorAtlas, learned magnetic embeddings) and discusses why this signal is not yet a default in VIO.
NetVLAD remains the load-bearing place-recognition baseline despite being seven years old. Successors — Patch-NetVLAD, MixVPR, MegaLoc, AnyLoc — improve recall on changing-condition benchmarks (day/night, summer/winter) but the gain narrows on the indoor case. The paper hypothesizes that interior featurelessness defeats the assumption that landmarks are stable. A ranked digest with per-method memory footprint and inference cost is included for practitioners.
Mobile depth estimation runs primarily on the NPU but degrades gracefully to CPU when the NPU is contended (camera ISP under load, video encode). The paper profiles MiDaS-Small, ZoeDepth-NK, and Depth Anything-Mobile under CPU-only conditions. Latency variance dominates; mean latency is the wrong metric.
Retrieval-augmented generation (RAG) has been adapted to spatial corpora. The paper reviews CLIP-feature retrieval, multi-vector retrieval (ColBERT-style), and hybrid sparse-dense retrieval over interior image stores. A finding worth filing: spatial retrieval benefits more from coarse layout features than from fine descriptors.
Mobile haptics — Taptic Engine, Android haptic API — provide a thin but useful feedback channel for AR confirmations. The paper reviews haptic-pattern design language (sharp vs. soft, single vs. paired) and reports user-detection thresholds at typical phone-holding postures.
Universal monocular-depth backbones (MiDaS, DPT, Depth Anything, ZoeDepth, Marigold) have moved from research curiosities to default scene-understanding components. The paper reports per-pixel scale-invariant error on indoor benchmarks and notes the backbone of choice now changes monthly. A short discussion of distillation paths to mobile is included.
Joint Embedding Predictive Architectures (JEPA) propose self-supervised learning by predicting in a learned representation rather than pixel space. V-JEPA extended this to video. The paper compares JEPA-class methods to MAE, DINOv2, and contrastive baselines on transfer to navigation and detection. Findings are mixed; the regime where JEPA wins is narrower than headline claims.
Differential privacy guarantees on spatial data are weaker than the literature implies. The paper reviews local-DP, central-DP, and shuffle-DP applied to room-scale telemetry. A finding: floor plans leak through low-ε mechanisms more than the worst-case theory suggests; ε ≤ 1 is required for meaningful protection.
Inside steel-framed buildings, GNSS pseudo-ranges drift faster than the IMU bias. The Kalman filter must be told to stop trusting the satellite stream. The paper reviews fault-detection-and-exclusion (FDE) heuristics, magnetometer cross-checks, and barometric altimetry as supplementary signals. A finding worth filing: pedestrian dead-reckoning (PDR) augmented by step-length estimation outperforms naive IMU integration by a factor of four on multi-floor traversals.
Most mobile cameras use rolling shutter; classical VIO assumes global shutter. The paper reviews per-row pose interpolation, learned RS undistortion (DeepRS, RS-NeRF), and the residual error after either. Findings: per-row interpolation suffices below 1 m/s motion; learned methods help on faster motion.
Differentiable rendering (Mitsuba 3, nvdiffrast, GAN-style implicit renderers) underpins most modern view synthesis. The paper indexes them by feature: physically-based, mesh, splat, neural. The lab uses these for ablation rather than runtime; runtime renderers are simpler.
Instance segmentation on mobile is dominated by YOLO-seg variants for cost reasons, but the gap to research-grade transformer methods (Mask2Former, MaskDINO) remains visible. The paper inventories where the gap matters (occluded thin structures) and where it does not (large convex objects).
INT8 quantization is now the default for mobile inference; FP16 remains common for newer NPU silicon with native FP16 paths. The paper reports per-op latency for representative kernels and notes that the crossover depends on the operator mix more than on the hardware.
Recovering a watertight mesh from a single moving camera is now feasible offline; doing so in real time on a phone is not. The paper reviews COLMAP-class structure-from-motion, Atlas-class learned reconstruction, NeuralRecon, MonoSDF, and SimpleRecon. Findings: learned methods are faster but hallucinate textureless walls; classical methods are slower but honest about uncertainty.
Sub-3B parameter VLMs are now within the inference budget of high-end mobile devices. The paper benchmarks Phi-Vision, MiniCPM-V, Idefics-Mobile, and Moondream on indoor caption-and-ground tasks. Findings: small VLMs are surprisingly capable on naming, mediocre on counting, and weak on spatial reasoning.
Before head-worn AR is universal, the phone is the pointer. The paper reviews ARKit/ARCore pointer abstractions, ray-casting from the phone center, and laser-pointer-style interaction. The dominant failure mode is fatigue under long sessions.
Most embodied-agent literature assumes a robot. A handheld device acting as a 'partial agent' (the human is the actuator) inverts the problem: perception is rich, motor primitives are gestures, success is user-confirmed. The paper sketches an evaluation protocol for handheld embodied tasks (find the door, follow the path, locate the tool).
Persistent anchoring across application launches is approached three ways in current systems: re-localize against a stored sparse map, re-localize against a learned descriptor cloud, or re-localize against a server-side spatial graph. Each pays a different price. Sparse maps are small and brittle to lighting; descriptor clouds are robust to lighting but heavy on storage; server graphs centralize what should be local. The paper enumerates the trade space and notes none of the three handles soft furnishings — the most common interior change.
Federated learning (FL) has been demonstrated at scale by major platforms (Gboard, Siri-style on-device updates). The paper reviews FedAvg, FedProx, SCAFFOLD, and client-drift mitigation. Findings: FL works for narrow tasks (next-word prediction, wake-word) and remains brittle for spatial tasks where client distributions differ wildly.
Synthetic objects without shadows look pasted on. The paper reviews learned shadow synthesis (Shadow-AR, ShadowGAN), classical shadow-casting from a single light estimate, and probe-based environmental capture. Findings: a single dominant light estimate suffices for most interior scenes.
SAM redefined what zero-shot segmentation could do. Its mobile descendants — MobileSAM, EfficientSAM, EdgeSAM — recover most of the quality at a fraction of the cost. The paper reports per-prompt latency and IoU on indoor classes. A short note: SAM-class models are *too* general for tracked-object pipelines; a tighter prompt budget is essential.
Neural Radiance Fields (NeRF) defined a generation of view-synthesis work. Five years on, the field has bifurcated: hash-grid encodings (Instant-NGP) for speed; tensor decompositions (TensoRF, K-Planes) for memory; and discrete primitives (3DGS) for both. The paper treats this as a Pareto front rather than a winner-takes-all, and indexes which method dominates which corner of the (training-time, render-time, memory) cube.
Grounding — answering 'where is X in this image?' — has matured rapidly. The paper compares CLIP-based zero-shot pointing, GroundingDINO box outputs, GLIP, KOSMOS-2 grounded captioning, and Molmo's pointing tokens. Findings: pointing-token methods are surprisingly accurate for a one-shot output but lossy under occlusion.
Sub-millisecond camera-IMU time alignment is a precondition for tight VIO. The paper reviews hardware-level synchronization, software estimation (Kalibr-style), and learned correction. A finding: published methods assume static alignment; in practice phones drift across thermal cycles.
Habitat and AI2-THOR have anchored embodied-agent research for half a decade. The paper benchmarks recurrent (LSTM-based) and transformer-based policies on PointGoal and ObjectNav, and reports a sim-to-real gap that has narrowed but not closed. Real-world deployment fails for reasons (clutter, thin obstacles, glass) that simulation under-models.
The lab's posture is that spatial data — room geometry, faces, voice — is processed on device unless the operator explicitly opts in to upload. The paper reviews the technical cost of this position (model size, compute envelope, storage) against the cost of the alternative (network egress, server-side retention, breach surface).
Across the visual-inertial odometry literature the headline accuracy figure is computed over sessions under three minutes. The lab finds error compounds non-linearly past the ten-minute mark in featureless interiors. The dominant failure mode is not gyroscope bias but loss of feature continuity during fast yaw; loop-closure rescues are rare without a place-recognition prior. The paper contrasts MSCKF, OKVIS, and VINS-Mono against an extended ORB-SLAM3 baseline on a 90-minute interior walkthrough.
Hands-free spatial UI is converging on a small vocabulary: gaze for targeting, pinch for selection, palm rotations for scaling, and voice for naming. The paper reviews input studies from major HMD platforms and notes the surprising consistency of pinch-as-select across vendors.
The mobile NPU landscape is a three-horse race in early 2025: Apple's Neural Engine, Qualcomm's Hexagon NPU, and Google's Tensor TPU. The paper benchmarks representative networks (YOLOv8n, MobileSAM, MiDaS-Small, Phi-3.5-mini) across each. Findings: the gap is narrower than vendor claims, wider than developer experience suggests.
The YOLO family has continued its quiet competence streak through versions 8–11. The paper benchmarks v8n, v9c, v10s, and v11n at 320×320 on three mobile NPU classes. Findings: v11n offers the strongest cost/accuracy ratio for interior-object detection; v8n remains the most predictable in latency variance. Not all gains generalize away from COCO.
VLMs (CLIP, BLIP-2, LLaVA, Qwen-VL, InternVL, Florence-2) have absorbed a generation of separate captioning, VQA, and grounding models. The paper inventories indoor-scene tasks where VLMs now dominate (object naming, attribute extraction, scene description) and tasks where they still underperform (spatial relations, counting, fine pose).
3D Gaussian Splatting (3DGS) reconstructs scenes as a cloud of anisotropic Gaussians and renders them at real-time rates on desktop GPUs. The paper inventories what is needed to deliver the same on mobile NPUs: a 4–10× reduction in splat count, INT8 quantization of opacity and SH coefficients, and a tiled rasterizer. The lab notes the open problem is not rendering but training: optimization currently demands 30+ minutes on consumer hardware.
AR rendering is fundamentally a compositing problem: the synthetic image must match the live image in exposure, color, and motion blur. The paper reviews real-time tone mapping, color-matching neural networks, and motion-blur transfer. The match is rarely perfect; user tolerance for the gap is wider than literature claims.
World models — agents that learn a latent forward dynamics model and plan inside it — have re-entered the mainstream after DreamerV3 demonstrated cross-domain generalization. The paper reviews the V1–V3 progression, IRIS, TWM, and the JEPA family. The transition from pixel-space to latent-space planning is treated as the field's central inflection.
IMU calibration splits into factory-stationary methods (Allan variance for noise characterization), in-situ stationary refinement, and online calibration during use. The paper reviews each and notes that online calibration has matured to where dedicated stationary calibration adds little for consumer devices. Allan-variance work remains essential for sensor selection.