Inertial Shape Estimation

• Inertial shape estimation is the process of reconstructing 3D geometry solely from inertial sensor data by mapping local acceleration and angular velocity cues to global shapes.
• It utilizes methods such as energy minimization, convex quadratic programming, and neural regression to resolve ambiguities and optimize shape parameters from IMU measurements.
• Applications span soft robotics, tensegrity mechanisms, and wearable motion capture, with experimental benchmarks demonstrating millimeter-level precision and real-time performance.

Inertial shape estimation is the process of recovering the three-dimensional geometry or configuration of a structure, mechanism, or body using measurements derived from inertial sensors—principally accelerometers and gyroscopes—in real time or offline. Unlike exteroceptive modalities (such as cameras, lidar, or structured light), methods in this category seek to reconstruct shape or anthropometric parameters solely or primarily from onboard inertial data. Originally driven by applications in soft robotics, tensegrity structures, and wearable motion capture, inertial shape estimation covers algorithmic, theoretical, and experimental frameworks for mapping local inertial cues to global geometric reconstructions in continuum robots, tensegrity manipulators, three-dimensional objects, and human bodies.

1. Mathematical Foundations and Model Classes

Inertial shape estimation is framed by a mathematical model relating sensor readings to the underlying geometric state. Central objects include:

• Continuum robot models: For soft and tensegrity robots, nodes and struts are represented as coordinates $\{\mathbf p_i\}$ and orientation unit vectors $\{\mathbf q_i\}$, with overall shape encoded in the spatial arrangement of elements. In class-1 tensegrity structures, node coordinates are given by

$$\mathbf n = \mathbf A^T \mathbf p + \mathbf B^T \mathbf q,$$

where $\mathbf A$ and $\mathbf B$ define geometric dependencies (Bhat et al., 16 Apr 2025, Bhat et al., 6 Mar 2026).

• Human body parameterizations: Anthropometric shape is encapsulated by vectors of principal shape coefficients ($\beta$), typically within the SMPL model framework. SMPL produces a mesh $M(\theta, \beta)$, where joint positions and surface locations are explicit functions of joint and shape parameters (Yin et al., 20 Oct 2025, Liu et al., 14 May 2025, Tang et al., 2 Mar 2026).
• Rigid body discretization: For inertial parameter identification, shapes are discretized into collections of points $\{^b p_i\}$, each assigned a nonnegative mass $m_i$, yielding a point-mass distribution that recovers the center of mass and inertia tensor (Nadeau et al., 2022).

The central challenge is to invert the mapping from local, generally ambiguous, inertial signals to the high-dimensional, redundant global shape variables, subject to physical, kinematic, or task-specific constraints.

2. Sensor-to-Shape Mapping in Physical Systems

The core of inertial shape estimation is the process of extracting global geometric state estimates from distributed or sparse measurements of acceleration and angular velocity.

• Inclination-based estimation (tensegrity/continuum robots): Individual six-axis IMUs yield local inclination angles $\phi_i$ by measuring the direction of gravity in the strut local frame, via:

$$\phi_i = \arccos\left(\frac{a_{iz}}{\|\mathbf a_i\|}\right).$$

The yaw $\theta_i$ (rotation about gravity) is not directly measured, leading to underdetermined orientation which is resolved by global optimization (Bhat et al., 16 Apr 2025, Bhat et al., 6 Mar 2026).

• Multipoint approach (vine robots): Absolute IMU orientation quaternions $q_i$ (possibly factory-calibrated) are referenced to a straight configuration to define relative orientations along the robot backbone. This forms a spatial kinematic chain where each joint/segment is described by relative hinge angles derived from sequential pairs of sensors (Laudenslager et al., 27 Feb 2026).
• Sparse body-worn IMUs (human pose and shape): Each IMU provides linear acceleration $a_i$ and angular velocity $\omega_i$ signals at specific mesh or joint locations. The mapping from $\{a_i, \omega_i\}$ to global pose and shape is confounded by inter-subject variations (limb length, girth, etc.), necessitating models that explicitly account for the combined influence of shape and pose (Yin et al., 20 Oct 2025, Liu et al., 14 May 2025, Tang et al., 2 Mar 2026).
• Point-mass discretization (object shape): Known object geometry, from CAD or vision, is discretized to recover spatially distributed inertial signals over points. This allows fusing shape priors with force/torque and motion data even when inertial cues are weak or noisy (Nadeau et al., 2022).

Sensor fusion and drift compensation, through algorithmic filters or probabilistic frameworks (e.g., complementary filters, Unscented Kalman Filter), are essential to mitigate the intrinsic integration errors and bias in raw inertial signals.

3. Optimization and Estimation Algorithms

Estimating shape from inertial measurements is operationalized as an optimization or inference problem, governed by physical constraints and sensor models.

• Energy minimization (tensegrity/soft robots): The physical principle is that the mechanically stable configuration minimizes the network's total elastic energy:

$$e(\mathbf p, \mathbf q) = \frac{1}{2} \sum_{k=1}^{m_s} K_k (m_k - b_k)^2, \quad m_k = \|\mathbf n_{i_k} - \mathbf n_{j_k}\|.$$

Optimizing over $\mathbf p$ and yaw angles $\boldsymbol\theta$ with measured inclinations $\boldsymbol\phi$, subject to geometric constraints (strut lengths, unit orientation vectors), yields node positions and global shape (Bhat et al., 16 Apr 2025, Bhat et al., 6 Mar 2026).

• Convex quadratic programming (inertial object parameters): Recovering spatial inertial parameters is cast as a quadratic program over nonnegative mass assignments:

$$\min_{\mathbf m \geq 0} \; \|(1-w)^\top(\widetilde A \mathbf m - \mathbf b)\|_2 + \|w^\top(A \mathbf m - \mathbf b)\|_2 + \lambda \|\mathbf m\|_2,$$

where $\widetilde A, A$ are regressors for reduced and full dynamics (Nadeau et al., 2022).

• Neural mapping and regression (human shape estimation): To disentangle pose and shape effects on IMU signals, decomposed regression models are trained to retarget and invert between template shape and individual-specific acceleration/velocity signals. Multilayer perceptrons (MLPs), recurrent neural networks, or spatiotemporal networks learn the correlation between shape code $\beta$, motion, and inertial signals, followed by physics-aware optimization coupling IMU consistency, dynamics, and anthropometric regularization (Yin et al., 20 Oct 2025, Liu et al., 14 May 2025, Tang et al., 2 Mar 2026).
• State-space probabilistic filtering (pose and shape fusion): For robust, drift-free estimation, tightly-coupled filtering, such as the Unscented Kalman Filter (UKF), fuses IMU, ultra-wideband (UWB) range, and learned pose predictions, enforcing kinematic consistency with shape priors and handling uncertainty due to drift or occlusion (Liu et al., 14 May 2025).

Gradient-based solvers, Gauss–Newton, Adam, and specialized QP solvers are commonly used, with convergence and real-time feasibility demonstrated across platforms.

4. Experimental Protocols and Quantitative Performance

Evaluation of inertial shape estimation encompasses hardware design, sensor placement, scenario diversity, and comparative results against ground truth.

• Tensegrity robots: Shape estimation on a Class 1 prism (4 struts, 12 springs, carbon fiber) yielded center-position mean absolute error (MAE) of 3.76 mm and node-position MAE of 20.78 mm after 300 gradient descent steps (≈97 ms); real-time operation exceeded 10 Hz. Yaw estimation was sensitive to actual yaw amplitude, with >80% error only in near-zero actual yaw cases (Bhat et al., 16 Apr 2025).
• Large-scale tensegrity manipulators: Five-layer, 20-strut structures, each equipped with an IMU, demonstrated length estimation accuracy of 2.1% (24.5 mm for 1160 mm true length) after 1000 optimization steps. Stability and convergence from arbitrary initial shapes, as well as robustness under manual tip perturbations, were observed (Bhat et al., 6 Mar 2026).
• Vine robots: Distributed IMUs (15–18 units) along a soft vine robot body enabled tip position estimation with mean errors of 11% (passive steering), 16% (active steering), and 8% across lengths of 30–175 cm. Orientation drift averaged 1.33°/min per IMU. Error scaled with length and IMU spacing—denser configurations improved spatial resolution but increased error from drift accumulation. Intermediate spacings (≈40–80 cm) performed best for simple curvatures (Laudenslager et al., 27 Feb 2026).
• Collaborative robot object identification: Point-mass discretization framework achieved mass errors ≤4.5%, center-of-mass errors 4.5–8.1%, and inertia error 29–37% in real-robot experiments; full pipeline run time was under 1 second (Nadeau et al., 2022).
• Human body motion and shape: Shape-aware inertial motion capture systems demonstrated mesh errors in shape estimation of 1.6–2.1 cm (T-pose), with joint-position error reduced by 19–23% vs. prior art across heights 110–190 cm. Joint space error for online fusion approaches was typically 5–6 cm, with mesh error under 6 cm. State-of-the-art pipelines extended this to metric-accurate real-time performance (>200 FPS), with positional errors <4 cm, and angular errors <10° for challenging datasets (Yin et al., 20 Oct 2025, Liu et al., 14 May 2025, Tang et al., 2 Mar 2026).

5. Error Sources, Limitations, and Physical Constraints

The primary sources of error and limitations in inertial shape estimation include:

• Sensor bias and drift: Low-frequency drift and bias in accelerometer and gyroscope signals limit the precision of inclination and angular velocity measurements, particularly in long-duration deployments (Bhat et al., 16 Apr 2025, Laudenslager et al., 27 Feb 2026).
• Ambiguity in unobservable axes: Absence of reliable magnetometer (yaw) data necessitates global optimization or fusion to resolve indeterminate orientations. For near-zero yaw displacements, optimization stalls or produces large errors (Bhat et al., 16 Apr 2025).
• Spatial discretization and underdetermined kinematics: Limited IMU count or sparse placement restricts the fidelity of curvature reconstruction and introduces tradeoffs between spatial resolution and cumulative sensor drift (Laudenslager et al., 27 Feb 2026).
• Generative model mismatch: For human shape estimation, conventional pose models trained on a single template underperform on nonstandard shape distributions (e.g., children, off-nominal body proportions), which motivates shape-aware correction networks and explicit shape-parameter estimation (Yin et al., 20 Oct 2025, Tang et al., 2 Mar 2026).
• Solver and computational efficiency: Larger or more complex structures require more sophisticated initialization, faster or higher-order solvers (Gauss–Newton, Levenberg–Marquardt), and real-time filtering pipelines (Bhat et al., 16 Apr 2025).

Physical validity (e.g., cable slackness, mechanical equilibrium) is often enforced as hard or soft constraints in energy minimization, QP, or filtering formulations.

6. Extensions, Opportunities, and Future Directions

Current research is expanding inertial shape estimation along several dimensions:

• Integration with exteroceptive sensing: Fusing inertial data with visual, range, or UWB sensors improves observability and error resilience, particularly for global translation and yaw axes (Liu et al., 14 May 2025, Tang et al., 2 Mar 2026).
• Advanced modeling and data-driven techniques: Learning-based mappings account for complex, nonparametric dependencies between shape, pose, and IMU signals, enabling generalization across body sizes and morphologies (Yin et al., 20 Oct 2025).
• Dynamic regime estimation: Embedding discrete-time or continuous-time process models (Lagrangian, state-space) allows extension to dynamic motions and offers compatibility with optimal filtering techniques (e.g., extended Kalman filter, smoothing) (Bhat et al., 6 Mar 2026).
• Scalability and multi-DOF structures: Stacking class-1 tensegrity modules and incorporating pressure-dependent stiffness, or adapting backbone parameterizations, allows extension to high-degree-of-freedom or continuum robots (Bhat et al., 6 Mar 2026).
• Hardware miniaturization and mechanical integration: Embedding smaller IMUs or distributed strain sensors offers increased deployment fidelity and reduced wiring overhead in robotic and wearable applications (Laudenslager et al., 27 Feb 2026).
• Physically consistent identification in collaborative robotics: The point-mass discretization approach guarantees positive semidefinite inertia recovery and compatibility with safety constraints in low-dynamics regimes for cobot manipulation (Nadeau et al., 2022).

Open challenges include further mitigating inertial drift, improving robustness to actuation and environmental variability, and extending real-time, joint pose–shape inference to new domains.