ReNiL: Relative Neural Inertial Locator

• ReNiL is a Bayesian deep-learning framework for pedestrian inertial localization, estimating 2D displacement and uncertainty from commodity IMU sequences.
• It uses a three-stage pipeline—motion-aware orientation filtering, an Any-Scale Laplace Estimator, and Bayesian inference—to align sensor data and predict displacement distributions.
• The demand-driven formulation with IPDPs enables adaptive waypoint spacing and robust fusion with external navigation sources for improved real-world tracking.

ReNiL denotes Relative Neural Inertial Locator, a Bayesian deep-learning framework for pedestrian inertial localization that estimates relative $2$D displacement and homogeneous Euclidean uncertainty from commodity IMU sequences, supports inference between Inertial Positioning Demand Points (IPDPs) at any scale, and links those estimates through a Bayesian inference chain (Wu et al., 8 Aug 2025). In current arXiv-adjacent usage, however, the same shorthand also appears as a user label for ReNeLiB in socially interactive agents, for RENI and RENI++ in inverse rendering, and, more loosely, for reversible nonlinear normalization layers in time-series forecasting; precise disambiguation is therefore essential.

1. Terminology and scope

Within the localization literature, ReNiL refers specifically to the framework introduced in "ReNiL: Relative Neural Inertial Locator with Any-Scale Bayesian Inference" (Wu et al., 8 Aug 2025). The same string is also used informally in several unrelated research contexts.

Usage	Expansion	Domain
ReNiL	Relative Neural Inertial Locator (Wu et al., 8 Aug 2025)	Pedestrian inertial localization
ReNiL (user shorthand)	ReNeLiB (Don et al., 2024)	Neural listening behavior for socially interactive agents
ReNiL / “RENI Lighting”	RENI / RENI++ (Gardner et al., 2023)	HDR natural illumination prior for inverse rendering
“ReNiL”-type methods	Reversible nonlinear normalization layers, exemplified by NoRIN (Zhang et al., 11 May 2026)	Time-series forecasting

This naming overlap is substantive rather than superficial. The four usages concern distinct technical problems: infrastructure-free inertial positioning, listener-behavior generation, illumination priors for inverse rendering, and reversible nonlinear normalization. A common misconception is that “ReNiL” names a single method family; in practice, the label is polysemous, and the full expansion or arXiv identifier is usually needed to identify the intended work.

2. Problem formulation in pedestrian inertial localization

ReNiL addresses pedestrian inertial localization using commodity IMUs on smartphones and IoT devices, with the goal of estimating relative position changes on a navigation plane without relying on WiFi, BLE, UWB, cameras, or other external infrastructure (Wu et al., 8 Aug 2025). The motivation is the persistent failure mode of classical INS and PDR pipelines under consumer MEMS noise, bias, and drift, together with the limitations of recent learning-based inertial odometry systems that rely on fixed sliding-window integration.

The framework isolates three limitations in prior learning-based methods. First, fixed sliding-window inference produces dense trajectories even when applications only require positions at specific times or events. Second, fixed-window models are inflexible with respect to motion scale and cadence, so their behavior changes when segment length changes. Third, predicted uncertainties are often inconsistent across scales and poorly suited to principled fusion with external sensing.

ReNiL reformulates localization around Inertial Positioning Demand Points:

$$T_d = \{t_0, t_1, \dots, t_n\}, \quad 0 \le t_0 < t_1 < \dots < t_n \le T.$$

These are time indices at which a position estimate is actually needed. Examples given for IPDPs include app queries, entering a semantic landmark such as a doorway or elevator, and periodic high-level application updates. Between adjacent IPDPs, the framework consumes the aligned IMU sequence

$$X^g_{t_n} = \{x^g_t\}_{t_{n-1} \le t \le t_n}, \quad x^g_t = [A^g_t, G^g_t] \in \mathbb{R}^6,$$

and predicts both displacement and uncertainty for that entire segment.

This demand-driven formulation changes the operational semantics of inertial localization. Rather than forcing every trajectory into a uniform, dense cadence, ReNiL permits application-controlled waypoint spacing. A plausible implication is that inference cadence can be matched to application semantics rather than imposed by network training conventions.

3. System architecture

ReNiL is organized as a three-stage pipeline: orientation alignment, Any-Scale Laplace Estimator (ASLE), and a Bayesian process over successive IPDPs (Wu et al., 8 Aug 2025). The architecture is explicitly probabilistic; the neural network does not merely output a point displacement but a displacement distribution parameterization.

Motion-aware orientation filter

The first stage aligns raw $9$-axis IMU data to a navigation frame. For device-frame accelerometer, gyroscope, and magnetometer measurements $A_t$, $G_t$, and $M_t$, and quaternion $q_t$, rotation is written as

$X^g = q^{-1} \otimes X \otimes q,$

and jointly as

$$\begin{pmatrix} A^g & G^g & M^g \end{pmatrix} = q^{-1} \otimes \begin{pmatrix} A & G & M \end{pmatrix} \otimes q.$$

The update is implemented as a motion-aware orientation filter based on gyroscope propagation with accelerometer and magnetometer correction through adaptive complementary fusion.

Two windows drive the adaptation. The accelerometer window $$T_d = \{t_0, t_1, \dots, t_n\}, \quad 0 \le t_0 < t_1 < \dots < t_n \le T.$$0 has length equal to one gait cycle, exploiting walking periodicity. The magnetometer window $$T_d = \{t_0, t_1, \dots, t_n\}, \quad 0 \le t_0 < t_1 < \dots < t_n \le T.$$1 is triggered only after the user has moved at least $$T_d = \{t_0, t_1, \dots, t_n\}, \quad 0 \le t_0 < t_1 < \dots < t_n \le T.$$2 meters, reducing sensitivity to local magnetic disturbances. The corresponding adaptive weights $$T_d = \{t_0, t_1, \dots, t_n\}, \quad 0 \le t_0 < t_1 < \dots < t_n \le T.$$3 and $$T_d = \{t_0, t_1, \dots, t_n\}, \quad 0 \le t_0 < t_1 < \dots < t_n \le T.$$4 depend on motion intensity, orientation stability, and magnetic consistency. This is a pedestrian-specific filtering strategy rather than a generic AHRS recipe.

Any-Scale Laplace Estimator

ASLE is the core network that maps an aligned IMU segment between two IPDPs to a Laplace-distributed planar displacement. The aligned accelerometer and gyroscope channels are concatenated into

$$T_d = \{t_0, t_1, \dots, t_n\}, \quad 0 \le t_0 < t_1 < \dots < t_n \le T.$$5

The sequence is split into temporal patches of length $$T_d = \{t_0, t_1, \dots, t_n\}, \quad 0 \le t_0 < t_1 < \dots < t_n \le T.$$6, with

$$T_d = \{t_0, t_1, \dots, t_n\}, \quad 0 \le t_0 < t_1 < \dots < t_n \le T.$$7

and tail padding applied when necessary.

The patching stage is followed by a TCN compressor that produces

$$T_d = \{t_0, t_1, \dots, t_n\}, \quad 0 \le t_0 < t_1 < \dots < t_n \le T.$$8

then a 1D ResNet feature extractor operating independently on patches, then a 2D ResNet contextual builder operating across the patch index and within-patch feature axis. The final PCO module performs multi-kernel adaptive pooling, concatenation, and regression to output

$$T_d = \{t_0, t_1, \dots, t_n\}, \quad 0 \le t_0 < t_1 < \dots < t_n \le T.$$9

Here $$X^g_{t_n} = \{x^g_t\}_{t_{n-1} \le t \le t_n}, \quad x^g_t = [A^g_t, G^g_t] \in \mathbb{R}^6,$$0 is the predicted displacement and $$X^g_{t_n} = \{x^g_t\}_{t_{n-1} \le t \le t_n}, \quad x^g_t = [A^g_t, G^g_t] \in \mathbb{R}^6,$$1 are Laplace scale parameters.

Several architectural constants are specified. The patch length is $$X^g_{t_n} = \{x^g_t\}_{t_{n-1} \le t \le t_n}, \quad x^g_t = [A^g_t, G^g_t] \in \mathbb{R}^6,$$2 samples at $$X^g_{t_n} = \{x^g_t\}_{t_{n-1} \le t \le t_n}, \quad x^g_t = [A^g_t, G^g_t] \in \mathbb{R}^6,$$3 Hz, corresponding to $$X^g_{t_n} = \{x^g_t\}_{t_{n-1} \le t \le t_n}, \quad x^g_t = [A^g_t, G^g_t] \in \mathbb{R}^6,$$4 s. The embedding module is a $$X^g_{t_n} = \{x^g_t\}_{t_{n-1} \le t \le t_n}, \quad x^g_t = [A^g_t, G^g_t] \in \mathbb{R}^6,$$5-layer TCN with embed channels $$X^g_{t_n} = \{x^g_t\}_{t_{n-1} \le t \le t_n}, \quad x^g_t = [A^g_t, G^g_t] \in \mathbb{R}^6,$$6, kernel size $$X^g_{t_n} = \{x^g_t\}_{t_{n-1} \le t \le t_n}, \quad x^g_t = [A^g_t, G^g_t] \in \mathbb{R}^6,$$7, stride $$X^g_{t_n} = \{x^g_t\}_{t_{n-1} \le t \le t_n}, \quad x^g_t = [A^g_t, G^g_t] \in \mathbb{R}^6,$$8, and GroupNorm size $$X^g_{t_n} = \{x^g_t\}_{t_{n-1} \le t \le t_n}, \quad x^g_t = [A^g_t, G^g_t] \in \mathbb{R}^6,$$9. The feature extractor has four residual layers with output channels $9$0, and the contextual builder has two residual layers with $9$1 channels. The final fully connected stage uses $9$2 hidden units, ReLU, and dropout. Because the final representation is obtained through adaptive pooling, ASLE accepts arbitrary segment length $9$3 and patch count $9$4.

Dual-task training

Training combines supervised Bayesian regression with self-supervised feature matching:

$9$5

The supervised term models displacement under a Laplace likelihood, while the self-supervised term enforces contextual feature consistency between original and augmented inputs. Augmentations include partial masking, quaternion constant bias interference, Gaussian noise, heading rotation, and abnormal protrusions. The stated purpose is robustness to sensor noise, orientation biases, and missing data.

The reported training setup uses PyTorch $9$6, a Xeon 4210R CPU and RTX 3090 GPU, Adam with learning rate $9$7, a ReduceLROnPlateau scheduler with factor $9$8, patience $9$9, minimum learning rate $A_t$0, batch size $A_t$1, and $A_t$2 training epochs.

4. Probabilistic model and any-scale Bayesian inference

A central feature of ReNiL is that each inter-IPDP displacement is modeled as a Laplace random variable rather than a deterministic vector:

$A_t$3

Per coordinate, the density is

$A_t$4

A common misconception is that ReNiL is only a point-estimation inertial odometry network. In the paper’s formulation, the network predicts both motion and uncertainty, and those uncertainties are propagated through a Bayesian chain (Wu et al., 8 Aug 2025).

For numerical stability, the model is trained by regressing average velocity over the IPDP segment rather than displacement directly. If $A_t$5 is the segment duration, then

$A_t$6

This makes the uncertainty scale with segment duration in a controlled way. The paper characterizes the resulting uncertainty as homogeneous Euclidean uncertainty: the predicted scales $A_t$7 and $A_t$8 are in meters and are directly interpretable as displacement-error scales.

The Bayesian chain uses a state transition of the form

$A_t$9

where $G_t$0 is the ASLE output and $G_t$1 is Laplace process noise. For pure inertial localization, successive IPDP posteriors are linked recursively. For fusion with external sources, the observation model is

$G_t$2

with Gaussian observation noise from the external sensor.

To incorporate Laplace process noise into a Kalman-like framework, ReNiL uses a Gaussian-exponential mixture representation, introducing latent variables $G_t$3 so that the conditional process model becomes Gaussian. The resulting inference procedure is described as a Rao–Blackwellized Kalman–Gibbs filter. This is the mechanism by which ReNiL supports loose fusion with GNSS, semantic landmarks, and related position observations.

5. Datasets, evaluation, and applications

The empirical evaluation uses RoNIN-ds and a new WUDataset (Wu et al., 8 Aug 2025). WUDataset was collected with a Google Pixel 6A at $G_t$4 Hz, with Vicon ground truth indoors at $G_t$5 Hz and LiDAR-SLAM outdoors at $G_t$6 Hz. It contains $G_t$7 participants, $G_t$8 trajectories, and more than $G_t$9 km of motion, with slow walking, normal walking, and running, and with multiple device placements including hand-held, shoulders, arms, front and back pockets, backpack, handbag, swinging arm, and chest pocket. Both WUDataset and RoNIN-ds use a train:validation:test split of $M_t$0, and the test split is further divided into seen and unseen subjects.

The baselines include PDR, RoRes18, TLIO, CTIN, iMoT, and IONet for FLOP comparisons. For direct comparison against fixed-window methods, the authors instantiate ASLE-1s, ASLE-5s, ASLE-10s, and ASLE-20s variants. The reported metrics are MAE, ADE, HE, QAE, and quaternion cosine similarity.

The orientation filter improves substantially over Mahony and Madgwick. On the Vicon subset, ReNiL reports QAE $M_t$1 versus $M_t$2 for Mahony and $M_t$3 for Madgwick, with cosine similarity $M_t$4 versus $M_t$5 and $M_t$6. On the LiDAR-SLAM subset, it reports QAE $M_t$7 versus $M_t$8 and $M_t$9, with cosine similarity $q_t$0 versus $q_t$1 and $q_t$2.

For displacement accuracy, the largest gains are reported on WUDataset. In the seen split, TLIO obtains MAE $q_t$3 m and ASLE-20s obtains $q_t$4 m. In the unseen split, TLIO obtains $q_t$5 m and ASLE-10s/20s obtain $q_t$6 m. On RoNIN-ds, improvements are smaller but still favorable: in the unseen split, TLIO reports MAE $q_t$7 m and the best ASLE variant reports $q_t$8 m. The paper interprets this as evidence that the any-scale architecture benefits more strongly as temporal scale increases and as motion becomes more complex.

The uncertainty analysis reports normalized-error coverage values of approximately $q_t$9, $X^g = q^{-1} \otimes X \otimes q,$0, and $X^g = q^{-1} \otimes X \otimes q,$1, and probability plots showing reasonable alignment for small-to-medium errors with larger deviations at high-error regimes. The stated interpretation is that the scale parameter retains consistent physical meaning across scales, which is the prerequisite for uncertainty ellipses and downstream fusion.

Efficiency results are also emphasized. For a $X^g = q^{-1} \otimes X \otimes q,$2 s inference, ReNiL / ASLE is reported at $X^g = q^{-1} \otimes X \otimes q,$3M parameters and $X^g = q^{-1} \otimes X \otimes q,$4G FLOPs, compared with RoRes18 at $X^g = q^{-1} \otimes X \otimes q,$5M and $X^g = q^{-1} \otimes X \otimes q,$6G, and TLIO at $X^g = q^{-1} \otimes X \otimes q,$7M and $X^g = q^{-1} \otimes X \otimes q,$8G. Despite having more parameters, ReNiL uses fewer FLOPs because of temporal compression in the embedding layer, parallel patch processing, and adaptive pooling.

The application studies are explicitly probabilistic. In behavioral semantic map construction, ReNiL is integrated with ATLAS-style semantic mapping and produces semantic landmark positions that are significantly closer to ground truth than those obtained from RoRes18. In a pure IPDP localization example, the orange dashed $X^g = q^{-1} \otimes X \otimes q,$9 uncertainty ellipse at each IPDP contains the ground-truth position throughout a long run. In fusion with behavioral semantic landmarks, the ReNiL process model is combined with semantic observations in the Rao–Blackwellized Kalman–Gibbs chain, again with ground truth remaining inside the reported $$\begin{pmatrix} A^g & G^g & M^g \end{pmatrix} = q^{-1} \otimes \begin{pmatrix} A & G & M \end{pmatrix} \otimes q.$$0 ellipses.

6. Limitations and broader naming landscape

The paper identifies several limitations of the localization framework (Wu et al., 8 Aug 2025). Calibration degrades somewhat for large errors and long temporal segments, which the authors attribute to relatively fewer long-segment examples. Cross-device and cross-activity generalization remain future-work targets, despite the reported robustness to unseen subjects and multiple device placements. The framework is also not yet positioned as a solution for ultra-low-power MCUs, and the demonstrated fusion examples are loosely coupled rather than tightly integrated with richer state augmentation or map constraints.

Outside inertial localization, the same string has distinct meanings. ReNeLiB, sometimes referred to by users as “ReNiL,” is a real-time software stack for neural listening behavior in socially interactive agents, combining multimodal cue extraction, a Learning-to-Listen-style behavior generator, and FLAME/ARKit avatar control (Don et al., 2024). RENI++, also described in the supplied material as a user shorthand target for “ReNiL,” is a rotation-equivariant, scale-invariant neural prior over HDR natural illumination for inverse rendering and environment completion (Gardner et al., 2023). In time-series forecasting, the NoRIN paper uses “ReNiL”-type language to denote the broader class of reversible nonlinear normalization layers, with NoRIN itself instantiated through a Johnson $$\begin{pmatrix} A^g & G^g & M^g \end{pmatrix} = q^{-1} \otimes \begin{pmatrix} A & G & M \end{pmatrix} \otimes q.$$1 transform and decoupled shape optimization (Zhang et al., 11 May 2026).

This multiplicity of usage suggests that “ReNiL” functions less as a stable field-wide term than as a context-sensitive shorthand. In localization, it denotes a specific Bayesian deep-learning framework centered on IPDPs, any-scale IMU inference, and Laplace uncertainty. In other domains, it refers to unrelated systems whose only commonality is the incidental reuse of the same label.