title: >- [Paper Note] Mitigating Radar-Inertial Calibration Ambiguities via SO(3) Manifold Steering description: >- [ECCV 2026][others][Spatiotemporal Calibration] A targetless Radar-inertial calibration framework that resolves mmWave angular ambiguities via temporal synthetic aperture and SO(3) spectral cycle-invariance. tags: - ECCV 2026 - others - Spatiotemporal Calibration - mmWave Radar - Inertial Measurement Units date: 2026-09-19 content_hash: b24fb8cde5a04c68
Mitigating Radar-Inertial Calibration Ambiguities via SO(3) Manifold Steering¶
Conference: ECCV 2026
Paper: ECCV Official
Code: https://github.com/MetaIoT-WHU/RadarCalib
Area: Autonomous Driving / 3D Vision
Keywords: mmWave Radar, Radar-Inertial Calibration, Temporal Synthetic Aperture, SO(3) Manifold Steering, Spectral Domain Alignment
TL;DR¶
RadarCalib introduces a targetless, plug-and-play Radar-inertial spatiotemporal calibration framework that circumvents point-matching ambiguities by extracting rotation-induced azimuthal drift in the spectral domain via temporal synthetic aperture (TSA), enforcing cycle-invariant manifold regularization alongside Doppler velocity and pose consistency on SE(3) to achieve sub-degree rotational and centimeter-level translation accuracy.
Background & Motivation¶
In adverse weather, low-illumination conditions, and feature-deprived operational domains, millimeter-wave (mmWave) Radar provides an indispensable sensing modality due to its weather immunity, penetrating capacity, and direct radial Doppler velocity measurements. However, fully realizing the potential of Radar-inertial fusion hinges upon precise spatial and temporal extrinsic calibration. Unlike cameras or LiDARs that yield dense geometric features and identifiable visual/structural primitives, mmWave Radar operates with a physically constrained antenna array, resulting in inherently coarse angular resolution (typically 10°–30°). This physical limitation, compounded by severe multipath effects, spurious reflections, and sparse, irregular point returns, renders conventional cross-frame geometric correspondence establishment highly unreliable.
Existing calibration paradigms typically depend on surveyed artificial targets such as trihedral corner reflectors or calibration boards, which precludes dynamic in-situ recalibration during autonomous robotic operation. Conversely, targetless motion-based techniques frequently impose restrictive kinematic assumptions—such as constant velocity or strictly static surroundings—that fail in realistic open-world environments. In feature-degraded regimes with sparse radar returns, point-based optimization landscapes lack sufficient topological constraints, frequently causing optimization to diverge, stall in degenerate local minima, or collapse completely on the SO(3) manifold.
This paper is motivated by a critical physical insight: while single-frame Radar point clouds are sparse and angularly ambiguous, Doppler velocity measurements remain exceptionally accurate, and 3D rotational motion continuously sweeps the antenna viewing direction over time, naturally synthesizing an extended temporal aperture (TSA). Furthermore, the frame-to-frame drift of the reconstructed spectral energy profile depends exclusively on the sensor's own rotation, remaining fundamentally invariant to external scene geometry and target distributions. The core idea is to synthesize a high-resolution azimuthal spectrum via rotational temporal aperture, extract a scene-agnostic cycle-invariant rotation constraint in the spectral domain, and tightly couple it with Doppler kinematics and short-horizon pose consistency for targetless joint SE(3) extrinsic optimization.
Method¶
Overall Architecture¶
RadarCalib accepts raw Radar ADC IF snapshots and synchronized high-rate IMU linear acceleration and angular velocity streams as inputs, and estimates the 6-DoF rigid transformation \(\mathbf{T}_I^R = [\mathbf{R}_I^R \mid \mathbf{t}_I^R] \in \mathrm{SE}(3)\) along with the temporal latency offset \(\tau\). The system first projects IMU-derived angular velocity onto the Radar antenna array's synthesis axis to drive a temporal synthetic aperture (TSA), recovering a high-resolution azimuthal energy spectrum; it then establishes a soft-aggregated \(2\pi\) cycle-invariance spectral residual that regularizes the rotation manifold against local minima. Concurrently, instantaneous Radar linear velocity is recovered from stationary Doppler returns via differentiable sinusoidal envelope fitting, creating a rigid-body velocity constraint through the kinematic lever-arm cross product. Finally, a short-horizon sliding-window pose consistency constraint bounds local trajectory drift. All three residual families are jointly solved via Lie-algebraic minimal updates in Google Ceres Solver.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Input Stream<br/>Radar ADC Snapshots + IMU Data"] --> B["Temporal Synthetic Aperture Enhancement<br/>Angular velocity virtual array sampling + coherent spectral beamforming"]
B --> C["Cycle-Invariant Spectral Regularization<br/>2π periodic soft-aggregation alignment for rotation residual"]
A --> D["Lever-Arm Doppler Velocity Constraint<br/>Sinusoidal Doppler fitting + rigid-body lever-arm kinematics"]
B & D --> E["Short-Horizon Pose Consistency Constraint<br/>Sliding-window trajectory alignment bounding local drift"]
C & D & E --> F["Joint SE(3) Manifold Optimization<br/>Two-stage Ceres optimization with minimal Lie algebra updates"]
F --> G["Calibration Output<br/>Accurate extrinsics T_I^R and time offset τ"]
Key Designs¶
1. Temporal Synthetic Aperture Enhancement: Overcoming Physical Array Resolution Limits
Single-frame automotive Radar returns suffer from severe azimuthal blur due to small antenna apertures, preventing reliable point feature matching. RadarCalib overcomes this physical barrier by synthesizing an extended antenna baseline over time through sensor rotation. Given the IMU angular velocity \(\boldsymbol{\omega}_I(t)\) and the array normal axis \(\mathbf{a}_R = \mathbf{e}_z\) in the Radar frame, the effective rotational rate driving the aperture synthesis is \(\omega_R(t) = \mathbf{a}_R^\top \mathbf{R}_I^R \boldsymbol{\omega}_I(t)\). For an array element located at \(\mathbf{p}_m = [d_r, 0, z_m]^\top\) (\(d_r\) being the effective rotation radius), continuous rotational motion generates a 3D cylindrical spatial sampling trajectory: $\(\mathbf{x}_m^W(t) = \begin{bmatrix} d_r \cos \phi(t) & d_r \sin \phi(t) & z_m \end{bmatrix}^\top, \quad \phi(t) = \phi_0 + \int_{t_0}^t \omega_R(\tau) d\tau\)$ Collecting complex snapshots \(s_m(t_k)\) across discrete sampling timestamps \(\{t_k\}_{k=0}^K\), the framework applies motion-adaptive weights \(w_k = \omega_R(t_k) / \bar{\omega}\) to compensate for nonuniform angular sampling caused by variable angular acceleration, and computes coherent far-field beamforming towards direction \(\mathbf{u}(\theta, \varphi)\): $\(y_t(\theta, \varphi) \propto \sum_{k=0}^K \sum_{m=1}^M w_k s_m(t_k) \exp\big(-j k\, \mathbf{u}(\theta, \varphi)^\top \mathbf{x}_m^W(t_k)\big)\)$ As accumulated rotation widens the synthetic baseline, the azimuthal mainlobe significantly narrows, transforming blurry spectral energy into sharp, resolvable angular signatures without modifying physical hardware.
2. Cycle-Invariant Spectral Regularization: Preventing Local Minima and Manifold Collapse
Marginalizing elevation \(\varphi \in \mathcal{B}\) yields a robust 1D azimuthal energy distribution \(E_t(\theta) = \sum_{\varphi \in \mathcal{B}} |y_t(\theta, \varphi)|^2\). Under pure self-rotation, adjacent profiles undergo an azimuthal translation \(E_{t+1}(\theta) \approx E_t(\theta + \omega_R(t)\Delta t)\). Directly regressing this differential inter-frame displacement is prone to noise-induced local minima. Instead, RadarCalib exploits the global algebraic property of \(2\pi\) spatial cycle invariance: completing a full \(2\pi\) yaw rotation \(\Psi(t) = \int_{t_0}^t \omega_R(\tau) d\tau\) returns the azimuthal spectral distribution to its origin. The framework builds a soft-aggregated reference spectrum using a Gaussian kernel centered at \(2\pi\): $\(\bar{E}_{t\circlearrowright}(\theta) = \sum_k \alpha_k(t) E_{t_k}(\theta), \quad \alpha_k(t) = \frac{\exp\big(-(\Psi(t_k) - \Psi(t) - 2\pi)^2 / (2\sigma_\kappa^2)\big)}{\sum_i \exp\big(-(\Psi(t_i) - \Psi(t) - 2\pi)^2 / (2\sigma_\kappa^2)\big)}\)$ The cycle-invariance spectral residual is formulated as a weighted log-power shape divergence: $\(\mathcal{R}_{\text{cycle}}(t; \mathbf{R}_I^R) = \mathbf{W}_\theta \big( \mathbf{N}(E_t) - \mathbf{N}(\bar{E}_{t\circlearrowright}) \big)\)$ where \(\mathbf{N}(\cdot)\) denotes log-power normalization and \(\mathbf{W}_\theta\) represents SNR-guided confidence weighting. This residual is fully differentiable with respect to \(\mathbf{R}_I^R\) and directly substitutes fragile kinematic angular rate matching, shrinking the feasible optimization space and eliminating convergence collapse under sparse observations.
3. Lever-Arm Doppler Velocity Constraint: Enabling Full Translation Observability
For stationary environmental reflectors, radial Doppler velocities follow a clean sinusoidal curve \(v_{ri} = -v_R(t) \cos(\alpha_i - \beta)\). Using differentiable least squares, the instantaneous linear velocity \(\mathbf{v}_R(t)\) is accurately extracted while dynamic moving scatterers are naturally filtered out as outliers. Through rigid-body kinematics, the Radar velocity maps to the IMU state via: $\(\mathbf{v}_R(t) = \mathbf{R}_I^R \big( \mathbf{v}_I(t) + \boldsymbol{\omega}_I(t) \times \mathbf{t}_R^I \big)\)$ where \(\mathbf{t}_R^I = -\mathbf{R}_I^R \mathbf{t}_I^R\) represents the translation lever arm. The cross product \(\boldsymbol{\omega}_I(t) \times \mathbf{t}_R^I\) renders the translational extrinsic vector \(\mathbf{t}_I^R\) fully observable under rotational excitation. For partially observable sensor orientations, an orthogonal projection matrix \(\mathbf{P}(t)\) isolates unobservable directions, producing the projected residual \(\tilde{\mathcal{R}}_v(t) = \mathbf{P}(t) \mathcal{R}_v(t)\).
4. Short-Horizon Pose Consistency Constraint: Stabilizing Local Trajectory Coherence
To eliminate integration drift and enforce global SE(3) structural integrity, the system couples short-horizon relative motion estimates: $\(\mathcal{R}_o(t) = \log\Big( \mathbf{T}_R(t)^{-1} \mathbf{T}_I^R \mathbf{T}_I(t) (\mathbf{T}_I^R)^{-1} \Big)^\vee\)$ Here, \(\mathbf{T}_I(t)\) is generated via standard preintegration with IMU bias and gravity compensation, and \(\mathbf{T}_R(t)\) is reconstructed using rotationally corrected Radar measurements. Evaluating this residual strictly over sliding windows of a few frames avoids long-term error accumulation and guarantees fast local convergence.
Loss & Training¶
The overall objective integrates all three residuals under Huber robust loss functions \(\rho(\cdot)\): $\(\min_{\mathbf{R}_I^R, \mathbf{t}_I^R} \sum_t \Big\{ \rho\big(\|\mathcal{R}_o(t)\|_{\boldsymbol{\Sigma}_o}^2\big) + \rho\big(\|\tilde{\mathcal{R}}_v(t)\|_{\boldsymbol{\Sigma}_v}^2\big) + \rho\big(\|\mathcal{R}_{\text{cycle}}(t)\|_{\boldsymbol{\Sigma}_\omega}^2\big) \Big\}\)$ Optimization is implemented in Google Ceres Solver using minimal Lie-algebraic updates: \(\mathbf{R}_I^R \leftarrow \mathbf{R}_I^R \exp(\widehat{\delta\boldsymbol{\theta}})\) and \(\mathbf{t}_I^R \leftarrow \mathbf{t}_I^R + \delta\mathbf{t}\). A two-stage optimization schedule is executed: Stage 1 refines \(\mathbf{R}_I^R\) using the spectral residual alone to clear large angular initialization errors; Stage 2 jointly optimizes both \(\mathbf{R}_I^R\) and \(\mathbf{t}_I^R\) across all terms. Solving the \(6 \times 6\) normal equation converges reliably within 30–45 iterations.
Key Experimental Results¶
Main Results¶
Real-world validation was conducted on a customized hardware platform featuring a TI AWR1843BOOST Radar and a MicroStrain 3DM-GX5-45 IMU mounted on a high-precision 3D-printed fixture with surveyed CAD ground-truth configurations (P.1 and P.2). RadarCalib is compared against the state-of-the-art targetless point-based calibrator iKalibr [8] across indoor Lab (L), Corridor (C), Hall (H), and Outdoor (O) sequences (rotation errors in degrees, translation errors in cm):
| Method | Scene & Placement | Rot. Error \(\delta\theta(r)\) | Rot. Error \(\delta\theta(p)\) | Rot. Error \(\delta\theta(y)\) | Trans. Error \(\delta p(x)\) | Trans. Error \(\delta p(y)\) | Trans. Error \(\delta p(z)\) | Time Offset \(\tau\) (ms) |
|---|---|---|---|---|---|---|---|---|
| RadarCalib (Ours) | Lab (L), P.1 | 0.11 ±0.04 | 0.22 ±0.22 | 0.15 ±0.25 | 0.03 ±0.12 | 0.21 ±0.17 | 0.05 ±0.03 | -68.27 ±0.13 |
| RadarCalib (Ours) | Corridor (C), P.1 | 0.25 ±0.22 | 0.14 ±0.21 | 0.12 ±0.12 | 0.21 ±0.23 | 0.02 ±0.14 | 0.03 ±0.15 | -57.19 ±0.16 |
| RadarCalib (Ours) | Corridor (C), P.2 | 0.20 ±0.22 | 0.15 ±0.17 | 0.18 ±0.19 | 0.28 ±0.22 | 0.19 ±0.18 | 0.25 ±0.17 | -57.85 ±0.28 |
| RadarCalib (Ours) | Hall (H), P.1 | 0.23 ±0.11 | 0.21 ±0.21 | 0.22 ±0.12 | 0.11 ±0.12 | 0.08 ±0.02 | 0.02 ±0.01 | -41.91 ±0.33 |
| RadarCalib (Ours) | Hall (H), P.2 | 0.25 ±0.24 | 0.18 ±0.20 | 0.21 ±0.22 | 0.22 ±0.10 | 0.16 ±0.12 | 0.19 ±0.20 | -56.60 ±0.42 |
| RadarCalib (Ours) | Outdoor (O), P.1 | 0.42 ±0.25 | 0.34 ±0.12 | 0.46 ±0.18 | 0.16 ±0.09 | 0.34 ±0.02 | 0.22 ±0.01 | -71.91 ±0.33 |
| RadarCalib (Ours) | Outdoor (O), P.2 | 0.40 ±0.35 | 0.42 ±0.28 | 0.35 ±0.26 | 0.30 ±0.10 | 0.15 ±0.12 | 0.18 ±0.14 | -56.60 ±0.12 |
| iKalibr [8] | Lab (L), P.1 | 0.61 ±0.96 | 0.57 ±0.82 | 0.63 ±0.80 | 0.23 ±0.31 | 0.35 ±0.29 | 0.23 ±0.18 | -66.60 ±0.17 |
| iKalibr [8] | Lab (L), P.2 | 0.73 ±0.64 | 0.75 ±0.72 | 0.88 ±0.83 | 0.68 ±0.94 | 0.67 ±0.88 | 0.61 ±0.79 | -76.45 ±0.21 |
| iKalibr [8] | Corridor (C), P.1 | 0.97 ±0.72 | 1.08 ±0.91 | 0.92 ±0.98 | 0.30 ±0.24 | 0.19 ±0.18 | 0.27 ±0.21 | -76.53 ±0.42 |
| iKalibr [8] | Corridor (C), P.2 | 1.02 ±0.47 | 0.80 ±0.51 | 0.83 ±1.07 | 0.75 ±0.80 | 0.58 ±0.92 | 0.63 ±0.72 | -76.25 ±0.34 |
| iKalibr [8] | Hall (H), P.1 | 0.60 ±0.72 | 1.02 ±0.88 | 0.95 ±0.90 | 0.36 ±0.30 | 0.34 ±0.27 | 0.39 ±0.31 | -86.60 ±0.45 |
| iKalibr [8] | Hall (H), P.2 | 0.81 ±1.16 | 0.90 ±1.01 | 1.08 ±0.98 | 0.78 ±0.66 | 0.62 ±0.61 | 0.81 ±0.75 | -76.40 ±0.25 |
| iKalibr [8] | Outdoor (O), P.1 | 1.25 ±0.72 | 1.02 ±0.88 | 0.95 ±0.90 | 0.36 ±0.30 | 0.34 ±0.27 | 0.39 ±0.31 | -76.53 ±0.46 |
| iKalibr [8] | Outdoor (O), P.2 | 1.41 ±1.36 | 1.30 ±1.11 | 1.28 ±0.98 | 0.78 ±0.66 | 0.62 ±0.61 | 0.81 ±0.75 | -86.62 ±0.14 |
Ablation Study¶
The contribution of each individual module was analyzed on real-world datasets:
| Config | Mean Rot. (deg) | \(\delta\theta(r)\) | \(\delta\theta(p)\) | \(\delta\theta(y)\) | Mean Trans. (cm) | \(X\) | \(Y\) | \(Z\) | Note |
|---|---|---|---|---|---|---|---|---|---|
| Full (Ours) | 0.254 | 0.264 | 0.245 | 0.254 | 0.173 | 0.121 | 0.189 | 0.210 | Full framework achieving sub-degree and mm-level accuracy |
| w/o TSA | 0.929 | 0.930 | 0.905 | 0.953 | 0.661 | 0.844 | 0.563 | 0.578 | Eliminating TSA increases rotation error by ~3.7× |
| w/o Velocity | 0.435 | 0.485 | 0.412 | 0.409 | 0.421 | 0.398 | 0.455 | 0.410 | Removing Doppler velocity degrades lever-arm observability |
| w/o Pose | 0.491 | 0.482 | 0.455 | 0.538 | 0.852 | 0.794 | 0.912 | 0.850 | Removing pose consistency increases translation error by ~5× |
Computational efficiency benchmarks (evaluated on Intel Xeon 6226R CPU @ 3.9 GHz): - RadarCalib (Ours): Converges in 45 iterations, 0.45 s/iteration, total time 20.25 s. - iKalibr [8]: Requires 120 iterations, 0.40 s/iteration, total time 48.00 s. - Speedup: RadarCalib achieves 2.4× end-to-end acceleration due to well-conditioned spectral constraints.
In downstream RIO SLAM benchmarks, RadarCalib reduces Absolute Trajectory Error (ATE RMSE) significantly across all scenes (Lab: 0.12 m vs 0.18 m; Corridor: 0.15 m vs 0.22 m; Hall: 0.18 m vs 0.25 m; Outdoor: 0.20 m vs 0.28 m), lowering global trajectory drift by up to 20%.
Key Findings¶
- TSA spectral alignment is the linchpin for resolving rotation ambiguity: Discarding TSA causes mean angular error to jump from 0.254° to 0.929°. Individual Radar frames with 10°–30° beamwidths are insufficient to constrain SO(3) rotations, whereas rotational aperture synthesis extracts unambiguous angular drift.
- Kinematic velocity and pose terms form a necessary complementary pair: Removing the pose consistency term causes translation error to degrade from 0.173 cm to 0.852 cm due to unmitigated integration drift, while removing the velocity term weakens the lever-arm cross product \(\boldsymbol{\omega}_I \times \mathbf{t}_R^I\), increasing translation error to 0.421 cm.
- Perceptual point cloud fidelity upgrade: Point clouds extracted from TSA-enhanced heatmaps decrease Chamfer Distance to 0.89–1.15 m (vs. 1.01–1.52 m for CFAR), elevate Valid Point Ratio (VPR) to 0.70–0.77 (vs. 0.56–0.70), and narrow the 3 dB beamwidth by 20%, resolving targets separated by just 2°–3°.
Highlights & Insights¶
- Turning dynamic rotation into an aperture synthesis asset: While conventional methods view rotational motion as a source of blur and distortion, RadarCalib leverages rotation to dynamically synthesize a large virtual antenna baseline, boosting azimuthal discriminability without hardware changes.
- Cycle-invariant spectral domain regularization: Replacing fragile point-to-point geometric correspondence matching with a continuous \(2\pi\) cyclic spectral shape divergence completely avoids local minima traps in sparse radar regimes.
- Lightweight Lie-algebraic optimization formulation: Formulating minimal updates on SE(3) requires solving only a \(6 \times 6\) linear system per iteration, converging in just 45 iterations and 20.25 seconds for practical in-situ deployment.
Limitations & Future Work¶
- Requirement for rotational excitation: Both the TSA mechanism and the translation lever-arm cross product rely on angular velocity (\(\boldsymbol{\omega}_I \neq \mathbf{0}\)). Under prolonged constant-velocity straight-line motion, rotational aperture synthesis cannot be fully activated.
- Horizontal 1D aperture limitation: Current automotive Radars utilize predominantly horizontal linear arrays, so TSA synthesis focuses on azimuth. Elevation angles are marginalized to reject clutter; future work could extend this to 3D full-aperture synthesis for 4D imaging radars.
- Online continuous recalibration: Integrating RadarCalib as an online factor within SLAM factor graphs will allow continuous tracking of slow extrinsic drifts induced by thermal expansion or chassis vibration.
Related Work & Insights¶
- vs iKalibr [8]: iKalibr uses continuous-time B-splines and point-level residuals, which frequently diverge or settle in sub-optimal local minima under sparse radar returns (yielding >1.0° rotation error). RadarCalib operates directly in the spectral domain, achieving sub-degree accuracy (0.254°) and 2.4× faster convergence.
- vs Target-Based Calibrators [10, 28]: Target-based setups require specialized trihedral corner reflectors and precise spatial surveying, preventing automated field deployment. RadarCalib is completely target-free and matches CAD-level precision.
- vs Learning-Based End-to-End Calibrators [24, 44]: Deep learning networks require massive training datasets and struggle to generalize across sensor placements or hardware shifts. RadarCalib relies purely on first-principles physics and differential geometry, requiring zero training data.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ [Pioneering use of temporal synthetic aperture and SO(3) cycle invariance for targetless radar-inertial calibration]
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Rigorous CAD-surveyed ground truth, multi-placement and multi-scene evaluations, ablations, runtime analysis, and SLAM verification]
- Writing Quality: ⭐⭐⭐⭐⭐ [Clear mathematical derivations, comprehensive physical explanations, and coherent narrative structure]
- Value: ⭐⭐⭐⭐⭐ [Provides an elegant, practical, and highly efficient solution to the longstanding radar angular resolution bottleneck]