Anchored Observation Models Explained

• Anchored observation models are a design principle where observations are interpreted relative to a fixed or internal reference, constraining inference with invariant structure.
• They are applied in diverse domains such as inertial navigation, visual estimation, generative reconstruction, and task planning to enhance system consistency and convergence.
• Empirical studies show that these models improve performance metrics like observability, reconstruction speed, and robustness under challenging observational conditions.

Searching arXiv for papers on anchored observation models and closely related anchored representations across navigation, vision, generation, and filtering. An anchored observation model is a formulation in which observations are interpreted relative to a fixed, persistent, or explicitly designated reference structure—an anchor—rather than as free-standing measurements. Across recent work, the anchor may be a fixed spatial point, a static scene region, an observed degraded signal, an anchor image, a symbolic-geometric world state, or an internal subset of salient variables. What unifies these formulations is that the anchor constrains inference or generation by tying it to an external reference or an explicitly structured internal representation. In navigation, a scalar range to a fixed anchor can be lifted into an augmented linear time-varying system (Sifour et al., 5 Dec 2025). In visual-inertial estimation, anchored landmark parameterizations alter the observability structure and improve consistency (Cohen et al., 17 Jun 2026). In generative modeling, observation-anchored flows and diffusion models condition the entire reverse process on a measured observation rather than pure noise (Wang et al., 2 Jul 2026, Wang et al., 21 Mar 2026). In fixed-view video grounding, static background anchors provide persistent semantic memory for re-entry and re-identification (Yan et al., 8 Mar 2026). More broadly, anchoring has also been used to reparameterize vision inputs relative to reference images (Narayanaswamy et al., 2024), to bind symbolic predicates to observable geometric anchors in mobile manipulation (Jiang et al., 28 Apr 2026), and to structure sequence generation around central anchor tokens or motifs (Klypa et al., 28 Feb 2025, Rout et al., 24 May 2025).

1. Anchoring as a General Modeling Principle

The term “anchor” is used in several technically distinct but structurally related senses. In single-range inertial navigation, the anchor is a fixed point $p_i^{\mathcal I}$ in the inertial frame, and the observation is the Euclidean distance to that point, $$d_y = \|p^{\mathcal I} - p_i^{\mathcal I}\| + \eta_{\text{uwb}}$$ (Sifour et al., 5 Dec 2025). In visual-inertial navigation, landmarks are parameterized relative to an anchor camera pose rather than directly in a global frame, so the measurement model depends on both the current pose and the anchor pose (Cohen et al., 17 Jun 2026). In fixed-view video grounding, an offline Anchor Bank is distilled from static background structures, and a query-conditioned Anchor Map becomes a persistent semantic memory over space (Yan et al., 8 Mar 2026).

In generative models, “observation-anchored” denotes a reverse process explicitly tied to the measured input. In Geo-Anchored Cloud Removal, the forward interpolant is defined as

$$x_t = \alpha_t x_* + \beta_t x_c + \sigma_t \epsilon,$$

so the trajectory is anchored to the cloudy observation $x_c$ rather than to pure noise (Wang et al., 2 Jul 2026). In mmWave-Diffusion, the forward chain injects the residual $\mathbf{z}=\mathbf{y}-\mathbf{x}$ between radar phase observation and clean respiration, and reverse sampling is initialized in an Observation-Consistent Neighborhood around $\mathbf{y}$ rather than from $\mathcal N(0,I)$ (Wang et al., 21 Mar 2026). In the Anchored Diffusion LLM, anchor variables are important tokens whose predicted distributions condition the denoising process through a two-stage reverse model (Rout et al., 24 May 2025).

A plausible implication is that anchored observation models are best understood not as a single algorithmic family but as a design pattern: they introduce a reference structure that persists across time, sampling steps, or abstraction levels, and they constrain estimation or generation relative to that structure. This suggests why anchoring appears in domains as different as robotics, remote sensing, diffusion modeling, and language modeling.

2. Fixed Spatial Anchors in Navigation and State Estimation

A canonical anchored observation model appears in single-range-aided inertial navigation. The system consists of an IMU, a single body-frame vector measurement such as a magnetometer, and one range measurement to a fixed anchor at known position $p_i^{\mathcal I}$ (Sifour et al., 5 Dec 2025). The anchored measurement is

$d = \|p^{\mathcal I} - p_i^{\mathcal I}\|,$

and, geometrically, each measurement constrains the position to a sphere centered at the anchor. The paper works with the squared range,

$y := \frac{1}{2} d_y^2,$

and shows that the scalar anchored distance can be embedded into an augmented linear time-varying system by introducing auxiliary quadratic coordinates $$d_y = \|p^{\mathcal I} - p_i^{\mathcal I}\| + \eta_{\text{uwb}}$$0, yielding a 13-dimensional extended LTV model with linear output $$d_y = \|p^{\mathcal I} - p_i^{\mathcal I}\| + \eta_{\text{uwb}}$$1 (Sifour et al., 5 Dec 2025). A Riccati/Luenberger-type observer on this system estimates body-frame position, velocity, and gravity direction; a complementary filter on $$d_y = \|p^{\mathcal I} - p_i^{\mathcal I}\| + \eta_{\text{uwb}}$$2 then reconstructs full orientation using the estimated gravity direction and a magnetometer vector. Under a uniform observability condition and non-collinearity of inertial gravity and magnetic-field vectors, the cascaded observer is Almost Globally Asymptotically Stable, with convergence to $$d_y = \|p^{\mathcal I} - p_i^{\mathcal I}\| + \eta_{\text{uwb}}$$3 and instability of the undesired $$d_y = \|p^{\mathcal I} - p_i^{\mathcal I}\| + \eta_{\text{uwb}}$$4-rotation equilibria (Sifour et al., 5 Dec 2025).

The observability condition is itself expressed in anchored terms. The augmented system is uniformly observable if there exist $$d_y = \|p^{\mathcal I} - p_i^{\mathcal I}\| + \eta_{\text{uwb}}$$5 such that

$$d_y = \|p^{\mathcal I} - p_i^{\mathcal I}\| + \eta_{\text{uwb}}$$6

with

$$d_y = \|p^{\mathcal I} - p_i^{\mathcal I}\| + \eta_{\text{uwb}}$$7

This requires sufficiently rich accelerations and rotations; degenerate motions can destroy observability even though the anchor is fixed and perfectly known (Sifour et al., 5 Dec 2025). The paper explicitly contrasts this minimal single-anchor setup with multiple-anchor range systems and richer range-bearing anchors, noting that the single-anchor case needs persistent excitation whereas multi-anchor systems are often directly solvable or observable under weaker conditions (Sifour et al., 5 Dec 2025).

An allied but distinct use of anchoring appears in filtering-based visual-inertial navigation. There, the anchor is not an environmental point but the camera pose relative to which a landmark is parameterized. For anchored inverse depth,

$$d_y = \|p^{\mathcal I} - p_i^{\mathcal I}\| + \eta_{\text{uwb}}$$8

and the global landmark position is recovered through the anchor pose and camera-IMU extrinsics (Cohen et al., 17 Jun 2026). The central observability result is that, with anchored feature parameterizations, the unobservable subspace is independent of the landmark state. For the landmark block, the nullspace contribution is

$$d_y = \|p^{\mathcal I} - p_i^{\mathcal I}\| + \eta_{\text{uwb}}$$9

whereas for globally parameterized landmarks the nullspace depends explicitly on the estimated feature position (Cohen et al., 17 Jun 2026). This improves estimator consistency because relinearizing the landmark state does not perturb the gauge structure. The paper further shows that all estimators employing anchored feature parameterizations exhibit improved consistency compared to estimators that resolve features in a global frame, especially under poor feature initialization, and that on TUM-VI the use of anchored feature representations alone can yield comparable performance to consistency-improved estimators employing a global feature representation (Cohen et al., 17 Jun 2026).

These two navigation examples use “anchor” differently—environmental reference point versus anchor pose—but both instantiate the same principle: a measurement model is reformulated around a reference entity whose presence changes observability and convergence properties.

3. Anchored Observation Models in Generative Reconstruction

In recent generative reconstruction work, anchoring is used to reshape both forward degradation models and reverse inference trajectories. In Geo-Anchored Cloud Removal, the Observation-Anchored Residual Flow defines the stochastic interpolant

$$x_t = \alpha_t x_* + \beta_t x_c + \sigma_t \epsilon,$$0

with a linear schedule $$x_t = \alpha_t x_* + \beta_t x_c + \sigma_t \epsilon,$$1, $$x_t = \alpha_t x_* + \beta_t x_c + \sigma_t \epsilon,$$2, $$x_t = \alpha_t x_* + \beta_t x_c + \sigma_t \epsilon,$$3 (Wang et al., 2 Jul 2026). The ideal velocity field is

$$x_t = \alpha_t x_* + \beta_t x_c + \sigma_t \epsilon,$$4

which explicitly decomposes into a clean target term, an observation-anchored term $$x_t = \alpha_t x_* + \beta_t x_c + \sigma_t \epsilon,$$5, and a stochastic term (Wang et al., 2 Jul 2026). The learned network $$x_t = \alpha_t x_* + \beta_t x_c + \sigma_t \epsilon,$$6 is trained by velocity matching, and inference integrates a deterministic ODE backward from a noisy observation-anchored terminal state. This anchored construction is paired with Geo-Contextual Prior Alignment, which constrains internal representations to remain close to a Vision Foundation Model feature manifold through the Geo-Contextual Integrity loss

$$x_t = \alpha_t x_* + \beta_t x_c + \sigma_t \epsilon,$$7

Empirically, OAR-Flow converges about $$x_t = \alpha_t x_* + \beta_t x_c + \sigma_t \epsilon,$$8 faster than EMRDM, and the full GACR about $$x_t = \alpha_t x_* + \beta_t x_c + \sigma_t \epsilon,$$9 faster, while improving reconstruction and downstream task performance across six cloud-removal datasets and twelve downstream tasks (Wang et al., 2 Jul 2026).

mmWave-Diffusion adopts an analogous but modality-specific construction for respiration sensing. The clean respiration waveform is $x_c$0, the radar phase observation is $x_c$1, and the residual is $x_c$2 (Wang et al., 21 Mar 2026). The forward kernel is

$x_c$3

with marginal

$x_c$4

so the forward chain explicitly follows the degradation path from respiration to radar observation (Wang et al., 21 Mar 2026). Reverse sampling starts from the Observation-Consistent Neighborhood

$x_c$5

and the Radar Diffusion Transformer is conditioned on $x_c$6 through dual positional encodings and banded-mask cross-attention (Wang et al., 21 Mar 2026). On 13.25 hours of synchronized radar-respiration data, the model achieves state-of-the-art waveform reconstruction and respiratory-rate estimation, with only 20 reverse steps at inference (Wang et al., 21 Mar 2026).

Anchoring also appears in discrete sequence diffusion. The Anchored Diffusion LLM identifies important tokens as anchors and introduces a two-stage reverse model: an anchor network predicts important-token distributions, and a denoiser predicts missing tokens conditioned on those anchored predictions (Rout et al., 24 May 2025). The corresponding Anchored Negative Evidence Lower Bound augments the diffusion objective with an explicit anchor prediction term,

$x_c$7

and the paper derives that

$x_c$8

Anchoring improves perplexity on LM1B and OpenWebText, with gains up to $x_c$9 over prior DLMs, state-of-the-art zero-shot generalization across seven benchmarks, and MAUVE scores surpassing autoregressive models (Rout et al., 24 May 2025).

A common misconception is that these models merely “add conditioning.” The papers instead define anchoring more strongly: the observation is built into the forward interpolant or forward corruption path, the reverse trajectory starts from an observation-consistent state, and the learned dynamics explicitly contain anchor-dependent terms (Wang et al., 2 Jul 2026, Wang et al., 21 Mar 2026, Rout et al., 24 May 2025).

4. Anchors as Persistent Memory, World-State Grounding, and Reference Frames

Anchored observation models are also used to maintain persistence across occlusion, absence, or world change. In fixed-view video grounding, AR$\mathbf{z}=\mathbf{y}-\mathbf{x}$0-4FV builds an Anchor Bank

$\mathbf{z}=\mathbf{y}-\mathbf{x}$1

from persistent background regions, where $\mathbf{z}=\mathbf{y}-\mathbf{x}$2 is a binary mask, $\mathbf{z}=\mathbf{y}-\mathbf{x}$3 an anchor prototype, and $\mathbf{z}=\mathbf{y}-\mathbf{x}$4 a centroid (Yan et al., 8 Mar 2026). A text query is aligned with these anchors by cosine similarity and temperature-scaled softmax to yield weights $\mathbf{z}=\mathbf{y}-\mathbf{x}$5, which define the Anchor Map

$\mathbf{z}=\mathbf{y}-\mathbf{x}$6

Because the camera is fixed and the anchors are static, $\mathbf{z}=\mathbf{y}-\mathbf{x}$7 is constant across time and functions as persistent semantic memory. When the target disappears, the system evolves an anchor-based re-entry prior $\mathbf{z}=\mathbf{y}-\mathbf{x}$8 by exponential moving average and Gaussian smoothing, and candidates are re-weighted using both the Anchor Map and $\mathbf{z}=\mathbf{y}-\mathbf{x}$9 (Yan et al., 8 Mar 2026). A ReID-Gating mechanism further combines appearance similarity, anchor evidence, and displacement in anchor coordinates. On long-term grounding in fixed-view videos, the system achieves a $\mathbf{y}$0 Re-Capture Rate improvement and a $\mathbf{y}$1 Re-Capture Latency reduction over the best baseline (Yan et al., 8 Mar 2026).

In home-service mobile manipulation, ANCHOR implements what the paper terms Physically Anchored Task Planning (Jiang et al., 28 Apr 2026). At each control cycle, a set of physical anchors $\mathbf{y}$2 is constructed from current sensor data, and symbolic state is derived deterministically as

$\mathbf{y}$3

Predicates such as near(r,o), aligned(r,o), holding(r,o), and in(o,c) are computed from geometric tests on anchors including robot pose, object pose, segmented point clouds, gripper current, and overlap ratios (Jiang et al., 28 Apr 2026). For example, near(r,o) holds if

$\mathbf{y}$4

and in(o,c) holds if

$\mathbf{y}$5

The planner reasons only over $\mathbf{y}$6, executes the first action of the current plan, then re-observes and recomputes anchors and predicates (Jiang et al., 28 Apr 2026). Across 60 real-robot trials, ANCHOR improves task success from $\mathbf{y}$7 to $\mathbf{y}$8 and achieves a $\mathbf{y}$9 recovery rate under perturbations (Jiang et al., 28 Apr 2026).

A closely related earlier formulation in semantic world modeling is perceptual object anchoring. There, an anchor $\mathcal N(0,I)$0 is an internal data structure that maintains the correspondence between a discrete symbol and perceptual sensor data over time and space (Persson et al., 2019). New percepts are matched to existing anchors using learned similarity scores over class, color, size, position, and time: $\mathcal N(0,I)$1 and a learned classifier decides whether to re-acquire an existing anchor or create a new one (Persson et al., 2019). With the time feature included, the best SVM achieves $\mathcal N(0,I)$2 accuracy and $\mathcal N(0,I)$3 F1 for the positive class (Persson et al., 2019). A high-level probabilistic tracker then maintains beliefs about occluded objects and feeds inferred positions back into the anchoring system, making the overall architecture a two-level anchored observation model that couples bottom-up association with top-down probabilistic reasoning (Persson et al., 2019).

These systems show that anchoring can function as persistent semantic memory, as explicit symbol grounding, or as a stable spatial reference frame. This suggests that anchored observation models are especially useful when long time horizons, disappearance, or symbolic-physical mismatch would otherwise cause drift.

5. Anchored Reparameterization and Internal Reference Structures

Not all anchored observation models rely on physical sensors or spatial landmarks. In vision training, anchoring is introduced as an architecture-agnostic reparameterization in which the network receives an anchor reference image $\mathcal N(0,I)$4 and residual $\mathcal N(0,I)$5, rather than the image $\mathcal N(0,I)$6 directly (Narayanaswamy et al., 2024). The anchored training objective is

$\mathcal N(0,I)$7

Prediction is intended to be invariant to the choice of $\mathcal N(0,I)$8, and inference may average over multiple references or use a single random reference (Narayanaswamy et al., 2024). The paper identifies a shortcut in vanilla anchored training: the model may ignore the reference and rely only on the residual. To counter this, it introduces a reference-masking regularizer that forces masked-reference tuples $\mathcal N(0,I)$9 to map to a near-uniform distribution. The resulting protocol improves OOD robustness, calibration, and anomaly detection across CIFAR and ImageNet benchmarks, with OOD accuracy gains up to about $p_i^{\mathcal I}$0 on ImageNet shifts and larger gains on CIFAR corruptions (Narayanaswamy et al., 2024).

Anchoring can also be imposed inside a sequence model’s generative order. In Bidirectional Anchored Generation for RNA design, two special tokens <ancl> and <ancr> are inserted into the sequence, and the model factorizes sequence probability by expanding outward from the anchor region: $p_i^{\mathcal I}$1 In practical RNA generation, anchors are placed inside binding-site regions, and a transformer with a custom BAnG attention mask learns to generate left and right context conditionally on this anchor region and a protein representation (Klypa et al., 28 Feb 2025). On synthetic motif tasks, BAnG achieves $p_i^{\mathcal I}$2 motif success on SingleBind and $p_i^{\mathcal I}$3 correct motif generation on DoubleBind, outperforming autoregressive and iterative baselines (Klypa et al., 28 Feb 2025). The same anchored conditional structure then transfers to protein-conditioned RNA design, where generated sequences outperform random sequences on DeepCLIP-based binding evaluation for 56 of 71 proteins (Klypa et al., 28 Feb 2025).

The Anchored Diffusion LLM extends this internal-reference idea to token importance rather than spatial position, while ACoT introduces explicit [ANT] tokens for reasoning, improving GSM8K to $p_i^{\mathcal I}$4 over a $p_i^{\mathcal I}$5 CoT baseline and reaching $p_i^{\mathcal I}$6 on ProntoQA (Rout et al., 24 May 2025). A plausible implication is that internal anchoring and external anchoring differ mainly in where the reference structure lives: in the environment, in the observation, or in the model’s factorization.

6. Constraints, Failure Modes, and Comparative Perspective

Anchored observation models typically improve robustness by restricting hypothesis space, but each formulation carries explicit assumptions. In single-range inertial navigation, the anchor location must be known and fixed, no sensor biases are estimated online, and uniform observability requires sufficiently rich accelerations and rotations; quasi-static or degenerate motions may be unobservable (Sifour et al., 5 Dec 2025). In visual-inertial navigation with anchored features, anchored parameterizations remove landmark-state dependence from the unobservable subspace, but the nullspace still depends on navigation state, so FEJ or RI-EKF remain useful on long trajectories (Cohen et al., 17 Jun 2026).

In generative reconstruction, the observation itself can become a limitation. OAR-Flow assumes that linear mixing of $p_i^{\mathcal I}$7 and $p_i^{\mathcal I}$8 in the interpolant is meaningful, which the paper notes may only approximate more complex radiative transfer effects (Wang et al., 2 Jul 2026). mmWave-Diffusion assumes near one-to-one temporal alignment between radar and respiration; if radar and reference signals are substantially misaligned, the dual positional alignment and banded masks may be less effective (Wang et al., 21 Mar 2026). ADLM depends on a definition of token importance, and the paper uses a simple frequency threshold rather than a semantically richer criterion (Rout et al., 24 May 2025).

Environment-anchored systems also rely on environmental stability. AR$p_i^{\mathcal I}$9-4FV assumes a fixed camera and stable background structures; moving cameras or strong background changes would invalidate the Anchor Bank (Yan et al., 8 Mar 2026). ANCHOR assumes that symbolic predicates can be re-derived from geometric evidence at each cycle; this reduces reliance on stale plans but requires continual successful perception and map maintenance (Jiang et al., 28 Apr 2026). In object anchoring, the anchoring system itself only maintains a single position per object in the permanent world model; multimodal location uncertainty is handled in the probabilistic tracker but not directly in the anchor representation (Persson et al., 2019).

Across domains, anchored formulations are often contrasted with unanchored or loosely conditioned alternatives. Single-range navigation is compared to multiple-anchor UWB localization and range-bearing systems (Sifour et al., 5 Dec 2025). OAR-Flow is contrasted with diffusion methods that start from pure noise and may hallucinate structure (Wang et al., 2 Jul 2026). mmWave-Diffusion is contrasted with DDPM-style sampling from $d = \|p^{\mathcal I} - p_i^{\mathcal I}\|,$0 (Wang et al., 21 Mar 2026). Anchored feature VINS is contrasted with global feature parameterizations (Cohen et al., 17 Jun 2026). Anchored vision training is contrasted with standard training on raw $d = \|p^{\mathcal I} - p_i^{\mathcal I}\|,$1 (Narayanaswamy et al., 2024). The common pattern is that anchoring shifts burden from unconstrained inference to structured conditioning.

A recurring misconception is that an anchor merely adds side information. The surveyed work instead treats the anchor as part of the model’s state, factorization, or observability structure. This suggests that the distinctive value of an anchored observation model lies not in extra data per se, but in the way a fixed or privileged reference reorganizes estimation or generation around invariant structure.

7. Synthesis and Research Directions

Across the literature, at least four recurring anchor types can be identified.

Anchor type	Representative role	Example papers
Fixed environmental reference	Provides sparse but persistent geometric constraint	(Sifour et al., 5 Dec 2025, Yan et al., 8 Mar 2026)
Anchor pose or parameterization frame	Re-expresses state to improve observability and consistency	(Cohen et al., 17 Jun 2026, Klypa et al., 28 Feb 2025)
Measured observation as anchor	Conditions forward and reverse generative processes on actual data	(Wang et al., 2 Jul 2026, Wang et al., 21 Mar 2026)
Internal salient/reference variable	Structures inference around important tokens or reference images	(Narayanaswamy et al., 2024, Rout et al., 24 May 2025)

This taxonomy suggests a unifying interpretation. An anchored observation model introduces a reference variable or structure $d = \|p^{\mathcal I} - p_i^{\mathcal I}\|,$2 such that inference is carried out not from observations alone but from observations relative to $d = \|p^{\mathcal I} - p_i^{\mathcal I}\|,$3, conditionals anchored on $d = \|p^{\mathcal I} - p_i^{\mathcal I}\|,$4, or states reparameterized by $d = \|p^{\mathcal I} - p_i^{\mathcal I}\|,$5. In robotics, $d = \|p^{\mathcal I} - p_i^{\mathcal I}\|,$6 may be a point, pose, or grounded object representation (Sifour et al., 5 Dec 2025, Cohen et al., 17 Jun 2026, Persson et al., 2019, Jiang et al., 28 Apr 2026). In perception and generation, $d = \|p^{\mathcal I} - p_i^{\mathcal I}\|,$7 may be a cloudy image, radar phase trace, token subset, or reference image (Wang et al., 2 Jul 2026, Wang et al., 21 Mar 2026, Rout et al., 24 May 2025, Narayanaswamy et al., 2024). In long-horizon video grounding, $d = \|p^{\mathcal I} - p_i^{\mathcal I}\|,$8 may be a query-aligned background memory (Yan et al., 8 Mar 2026).

Several future directions are explicitly identified in the source works. Single-range navigation points to multiple anchors, bias estimation, and discrete-time observers (Sifour et al., 5 Dec 2025). GACR suggests multi-temporal and multi-modal anchored flows, more realistic physical forward models, and semi-supervised or unpaired anchoring (Wang et al., 2 Jul 2026). mmWave-Diffusion suggests extending residual-based anchored diffusion to other inverse problems such as medical imaging, audio denoising, and geophysical inversion (Wang et al., 21 Mar 2026). Anchored feature VINS points to consistency analysis for optimization-based VINS and broader SLAM settings (Cohen et al., 17 Jun 2026). BAnG suggests transfer to other domains where localized functional subsequences govern global structure (Klypa et al., 28 Feb 2025). Anchored vision training suggests that anchoring is a representation prior rather than merely a training trick, and may be combined with strong augmentation and uncertainty estimation protocols (Narayanaswamy et al., 2024).

Taken together, the literature indicates that anchored observation models are most effective when an external or internal reference can be treated as more stable, informative, or semantically privileged than the rest of the observation stream. Their practical effect is to convert underconstrained inference into constrained inference; their theoretical effect is often to expose invariant structure, alter observability, or reduce effective sample complexity.