Kinematic Observer & Predictor

• Kinematic Observer and Predictor is an estimation framework that reconstructs unmeasured motion variables using a propagation law combined with measurement corrections.
• The approach incorporates geometric formulations on manifolds, delay compensations, and learned latent-state models to handle intermittent and noisy sensor data.
• It offers practical insights for robust state estimation in robotics and vehicles by integrating kinematic information with dynamic internal models.

A kinematic observer and predictor is an estimation architecture that reconstructs unmeasured motion variables from partial measurements while propagating an internal state through a kinematic or geometry-preserving model. In the literature summarized here, this architecture appears in several forms: nonlinear observers on manifolds such as $SO(3)$, $SE(3)$, $S^2$, and $T\mathbf G$; predictor-based compensators for delayed sensing; vehicle and legged-robot estimators that combine kinematics with interaction dynamics; and learned latent-state models that infer hidden state from measurement histories and then roll it forward under candidate inputs (Ng et al., 2021, Senejohnny et al., 2016).

1. Core architecture and recurring design pattern

Across otherwise different domains, the recurring structure is a propagation law plus a correction law. In discrete-time attitude estimation on $\SO3$, the predictor is an exponential propagation step,

$$\hat R_d(k\mid k-1)=\hat R_d(k-1\mid k-1)\exp(\hat\Omega_d(k-1)_\times\Delta t),$$

and the corrector injects innovation and an off-manifold recovery term weighted by $k_e$, producing a genuine predictor-corrector observer rather than a purely algebraic update (Shanbhag et al., 2019). In missile guidance, the same architecture appears at signal level rather than state-manifold level: a first stage estimates current derivatives of a delayed seeker signal $v(t)$, and a second stage estimates the derivatives at $t+\Delta$, with asymptotic objective

$$x_{i,2}\to v^{(i-1)}(t+\Delta),\qquad i=\{1,2,3,4\},$$

so that the predicted future LOS-rate can be fed directly to the guidance law (Smithson et al., 2024). In learned locomotion planning, the observer reconstructs a latent state $SE(3)$0 from a history of proprioceptive measurements and commands, and a separate predictor rolls that latent state forward under many candidate future command sequences for MPPI (Kulkarni et al., 26 Oct 2025).

This suggests that “predictor” is often not a separate theorem or standalone module. In some papers it is the explicit time-update map of a predictor-corrector observer; in others it is the internal model that evolves between measurement corrections; and in still others it is a delay-compensating extrapolator acting directly on measured signals.

2. Geometric formulations on manifolds and Lie groups

A major strand of the literature treats kinematic observation as an intrinsically geometric problem. For second-order kinematic systems on a matrix Lie group $SE(3)$1, the natural state is not $SE(3)$2 alone but the tangent bundle $SE(3)$3, with state $SE(3)$4. The “classical” second-order kinematics are

$SE(3)$5

with input space $SE(3)$6. The symmetry-consistent construction uses the semi-direct product $SE(3)$7, a right action on state and input, and an equivariant lift

$SE(3)$8

A central point is that the virtual first-order input $SE(3)$9 is “critical for the development of the equivariance properties of second order systems.” The observer is then posed on the lifted group, and the formal stability result is local asymptotic stability of the error equilibrium under boundedness assumptions on $S^2$0 and $S^2$1 (Ng et al., 2021).

For rigid-body rotation on $S^2$2, the geometric issue is different: numerical discretization of continuous-time kinematics generally violates $S^2$3 if standard Euclidean schemes are applied naively. The modified Mahony observer adds the feedback-integrator term

$S^2$4

which vanishes on $S^2$5 but makes the manifold attractive off-manifold. The resulting discrete-time observer retains a predictor-corrector form, handles constant gyro bias, and converges to the modified continuous-time observer as $S^2$6; for sufficiently small $S^2$7, the paper concludes local exponential convergence (Shanbhag et al., 2019).

On $S^2$8, pose-only estimation for a free-floating rigid body introduces a different geometric difficulty: velocity lives in a distinct tangent space and must be reconstructed without direct measurement. The observer therefore estimates both target pose $S^2$9 and body velocity $T\mathbf G$0, using a logarithm-based pose innovation and an internal model of free rigid-body dynamics. The resulting error system is shown to be almost-globally uniformly asymptotically stable, and because the observer embeds the rigid-body dynamics, it may act as a predictor during periods of slow, noisy, or missing pose measurements (Mishra et al., 2019).

Equivariant observer design also extends to visual localization and mapping and to bearing dynamics. In visual SLAM, the state space is a quotient manifold built from robot pose on $T\mathbf G$1 and landmarks in $T\mathbf G$2, allowing finite points and pure bearings to share one representation. The observer is posed on a symmetry group, uses decoupled $T\mathbf G$3 Riccati gains for point landmarks, and is shown to be almost globally asymptotically stable and exponentially stable in-the-large on the SLAM manifold (Goor et al., 2019). For first-order bearing dynamics on $T\mathbf G$4, an equivariant lift to $T\mathbf G$5 yields an observer for a model with both angular velocity and an additional linear-velocity-driven tangent term, with projected form

$T\mathbf G$6

and almost global asymptotic stability (Serrano et al., 16 Dec 2025).

3. Delay compensation, sparse measurements, and explicit prediction

A second major strand treats prediction as compensation for delayed or intermittent measurements. For planar rigid-body motion in $T\mathbf G$7 with delayed landmark outputs, the state is embedded in Euclidean coordinates as

$T\mathbf G$8

and the delay is modeled as a transport PDE. The predictive observer is first formulated as an ODE–PDE system and then rewritten in a PDE-free realization with an integral term that transports innovation over the delay interval. The error system is proved globally exponentially stable when the constant known delay satisfies $T\mathbf G$9, the gain parameter lies in an admissible interval $\SO3$0, and the landmark configuration makes $\SO3$1 full rank; in planar terms, the landmarks must not all lie on one straight line (Senejohnny et al., 2016).

In missile guidance, the predictor is explicit and signal-centered. The observer receives a delayed LOS-rate measurement $\SO3$2, estimates current derivatives through a high-gain differentiator cascade, and then estimates future derivatives through $\SO3$3-weighted gains. The predicted future LOS-rate $\SO3$4 replaces the delayed measurement in proportional navigation. The reported non-manoeuvring-target result is a reduction of miss distance from $\SO3$5 to $\SO3$6, with LOS-rate RMSE reduced from $\SO3$7 to $\SO3$8; the Monte Carlo study further indicates useful compensation across delays from $\SO3$9 to $$\hat R_d(k\mid k-1)=\hat R_d(k-1\mid k-1)\exp(\hat\Omega_d(k-1)_\times\Delta t),$$0, with near-flat corrected miss distance up to about $$\hat R_d(k\mid k-1)=\hat R_d(k-1\mid k-1)\exp(\hat\Omega_d(k-1)_\times\Delta t),$$1 delay (Smithson et al., 2024).

Cloud vehicle control under bi-directional time-varying delay uses a different predictor-observer construction. The design model is an uncertain delayed linear system with measurable output delay $$\hat R_d(k\mid k-1)=\hat R_d(k-1\mid k-1)\exp(\hat\Omega_d(k-1)_\times\Delta t),$$2 and bounded but unmeasurable input delay $$\hat R_d(k\mid k-1)=\hat R_d(k-1\mid k-1)\exp(\hat\Omega_d(k-1)_\times\Delta t),$$3. The key device is the transformed state

$$\hat R_d(k\mid k-1)=\hat R_d(k-1\mid k-1)\exp(\hat\Omega_d(k-1)_\times\Delta t),$$4

together with an output reconstruction $$\hat R_d(k\mid k-1)=\hat R_d(k-1\mid k-1)\exp(\hat\Omega_d(k-1)_\times\Delta t),$$5 built from $$\hat R_d(k\mid k-1)=\hat R_d(k-1\mid k-1)\exp(\hat\Omega_d(k-1)_\times\Delta t),$$6, known past controls, and the delay bounds. The observer update

$$\hat R_d(k\mid k-1)=\hat R_d(k-1\mid k-1)\exp(\hat\Omega_d(k-1)_\times\Delta t),$$7

is then embedded in a delay-free interconnected system used for robust asymptotic stability analysis. This is not a purely kinematic observer in the narrow sense, but it is a canonical predictor-observer architecture for delayed state reconstruction (Pan et al., 2023).

The same theme appears more implicitly in free-floating target observation. There the observer is not marketed as a delay compensator, but because it contains a full internal model for pose and body velocity, it may act as a predictor during periods of slow-sampled or missing pose measurements (Mishra et al., 2019). This suggests that intermittent-sensing prediction and explicit delay compensation are closely related design motifs.

4. Vehicles, contact-rich locomotion, and mixed kinematic–dynamic observers

In many practical systems, direct measurement of the desired variable is unavailable, and the resulting observer is only partly kinematic. A clear example is lateral velocity estimation for vehicles. The adaptive sliding mode observer is built on a 2-DOF bicycle model, but it exploits the measurable derivative relation

$$\hat R_d(k\mid k-1)=\hat R_d(k-1\mid k-1)\exp(\hat\Omega_d(k-1)_\times\Delta t),$$8

to construct a filtered virtual measurement of the unmeasured lateral velocity $$\hat R_d(k\mid k-1)=\hat R_d(k-1\mid k-1)\exp(\hat\Omega_d(k-1)_\times\Delta t),$$9. The method is therefore a dynamic observer augmented by derivative-based kinematic information rather than a purely kinematic estimator. Its most emphasized test is the slow J-turn on low-$k_e$0, where the SMO tracked the optical measurement nicely while the Farrelly & Wellstead kinematics approach underestimated rear slip angle by about $k_e$1, or roughly $k_e$2 deg, at a point where the actual value was about $k_e$3 deg (Tseng, 27 Jun 2026).

For legged robots, the “Kinetics Observer” is a tightly coupled MEKF that simultaneously estimates contact and perturbation forces and the robot’s kinematics. Its predictive model is driven by centroidal Newton–Euler dynamics, while its corrections use IMU signals, encoder-derived kinematics, and contact wrench measurements linked by a visco-elastic contact model. The coupling produces measurement redundancy and is reported to improve robustness and accuracy. On HRP-5P, the final position error is $k_e$4 versus $k_e$5 for legged odometry, and the final orientation error is $k_e$6 versus $k_e$7; runtime is under $k_e$8 per iteration on a laptop CPU (Demont et al., 2024).

Motorcycle state estimation pushes the same mixed viewpoint toward hazard-oriented state reconstruction. The four-rigid-body model has seven degrees of freedom and augments the state with tire-force dynamics, yielding

$k_e$9

A Luenberger observer based on LQR theory is then designed from IMU measurements $v(t)$0. The stated purpose is to estimate hidden variables useful for predicting dangerous scenarios such as lowside, highside, and fall. In rectilinear tests against BikeSim, the reported absolute static error of longitudinal tire forces is less than $v(t)$1, and in off-nominal speed tests an observer designed at $v(t)$2 yields $v(t)$3 static error at $v(t)$4 and $v(t)$5 at $v(t)$6 (Kabwangala et al., 2024).

These works suggest that, in applied robotics and vehicles, “kinematic observer” often denotes an estimator whose outputs are kinematic variables of interest even when the internal model includes contact mechanics, tire relaxation, or centroidal dynamics.

5. Learning-based observers and neural predictor models

A newer strand replaces explicitly derived state coordinates by learned latent representations while retaining the observer–predictor decomposition. On the kinematic bicycle benchmark, the learning-based observer maps a recent history of noisy measurements $v(t)$7 and known controls $v(t)$8 to the current latent state $v(t)$9. The benchmark uses the discrete-time model

$t+\Delta$0

with $t+\Delta$1. At nominal noise $t+\Delta$2, EKF is best; at $t+\Delta$3, the reported RMSEs are $t+\Delta$4 for EKF versus $t+\Delta$5 for the best CNN on $t+\Delta$6, and the CNN overtakes EKF around $t+\Delta$7 (Ghosn et al., 2023).

In cybersickness prediction, the observer notion is broadened further. The final deployed model uses only kinematic inputs at inference time, but it is trained to infer a latent physiological embedding that matches an EDA-derived representation of susceptibility. The training objective is

$t+\Delta$8

where $t+\Delta$9 is MSE or MAE between the kinematically inferred physiological embedding and the EDA-encoded representation. The best reported result is $$x_{i,2}\to v^{(i-1)}(t+\Delta),\qquad i=\{1,2,3,4\},$$0 accuracy for 4-class cybersickness prediction using 10 seconds of historical kinematic data only at inference, with physiological data used only during training as privileged supervision (Li et al., 2023). This suggests that learned “kinematic observers” may estimate hidden internal-state surrogates rather than only physical coordinates.

For legged locomotion planning, the learned observer–predictor architecture makes this decomposition explicit. The observer is

$$x_{i,2}\to v^{(i-1)}(t+\Delta),\qquad i=\{1,2,3,4\},$$1

with $$x_{i,2}\to v^{(i-1)}(t+\Delta),\qquad i=\{1,2,3,4\},$$2, and Theorem 1 states that if

$$x_{i,2}\to v^{(i-1)}(t+\Delta),\qquad i=\{1,2,3,4\},$$3

then the state and output estimation errors are uniformly ultimately bounded. The predictor is a GRU rollout

$$x_{i,2}\to v^{(i-1)}(t+\Delta),\qquad i=\{1,2,3,4\},$$4

used to evaluate thousands of candidate trajectories for MPPI. The reported runtime is $$x_{i,2}\to v^{(i-1)}(t+\Delta),\qquad i=\{1,2,3,4\},$$5 for $$x_{i,2}\to v^{(i-1)}(t+\Delta),\qquad i=\{1,2,3,4\},$$6 trajectories over $$x_{i,2}\to v^{(i-1)}(t+\Delta),\qquad i=\{1,2,3,4\},$$7 steps on an NVIDIA RTX T1000, and hardware experiments reduce failure rate from $$x_{i,2}\to v^{(i-1)}(t+\Delta),\qquad i=\{1,2,3,4\},$$8 for the constant-velocity baseline to $$x_{i,2}\to v^{(i-1)}(t+\Delta),\qquad i=\{1,2,3,4\},$$9 for the proposed method (Kulkarni et al., 26 Oct 2025).

6. Broader meanings, misconceptions, and conceptual limits

A distinct use of the term appears in relativity, where “observer” denotes a physical observer field rather than an estimation algorithm. In Schwarzschild spacetime, the kinematic relative velocity of a test particle with respect to a stationary observer is defined by parallel transport and spacelike simultaneity,

$SE(3)$00

For stationary observers, the modulus satisfies

$SE(3)$01

so it depends on the test particle event and $SE(3)$02, but not on which stationary observer is chosen. For circular geodesics, $SE(3)$03, depending only on orbital radius (Bolós, 2012). This is not a control observer, but it is an observer-dependent kinematic comparison law.

A related frame-dependent use appears in gravitational-wave propagation. Under a boost, the observed frequency transforms by the Doppler factor, the arrival direction is aberrated, and the polarization tensor rotates with the same rotation that acts on the wave direction. The crucial asymmetry is that source motion changes the strain amplitude by the Doppler factor, whereas observer motion leaves the amplitude invariant; by contrast, specific intensity transforms in the same way for source and observer boosts (Cusin et al., 2024). Here again, “observer” refers to a physical frame of measurement rather than a state estimator.

Taken together, the literature suggests several clarifications. A kinematic observer is not necessarily a Euclidean additive filter; geometric formulations instead preserve manifold structure and often rely on equivariance or multiplicative error. A predictor is not always a separate algorithm; in many observer designs it is simply the internal model used between corrections. “Kinematic” does not necessarily exclude dynamics, since many practical observers estimate kinematic outputs through dynamic internal models of contact, tire forces, or closed-loop actuation. Finally, the term “observer” can denote either an estimation algorithm or a physical frame, and the intended meaning depends strongly on context.