acceptodds
Under review as a conference paper at ICLR 2027

Observability-Aware Manifold Flow Matching for Degradation-Resilient Radar-Inertial-Visual UAV Odometry

Abstract

Reliable metric odometry for small unmanned aerial vehicles is difficult when vision is degraded and radar returns have limited directional coverage. We present Observability-Aware Manifold Flow Matching (OAM-FM), a self-supervised radar-inertial-visual odometry framework that learns conditional transport in local pose coordinates and maps its outputs to SE(3). A calibrated Doppler Jacobian supplies a directional training metric; its undamped information matrix characterizes instantaneous radar constraints without conflating numerical regularization with physical observability. The objective combines flow matching, robust Doppler residuals, and IMU preintegration over poses, velocities, and sensor biases. Optional visual features provide complementary constraints without being required at inference. A separate covariance head estimates local prediction uncertainty, whose calibration is evaluated against reference poses. The formulation distinguishes this uncertainty from the spread of flow samples: a deterministic sensor-based training endpoint alone does not establish a multimodal posterior. The method requires no dense depth or pose labels during training and is evaluated on real-world UAV odometry under radar sparsity, visual degradation, aggressive motion, and inertial bias, together with onboard latency.

Primary area:
applications to robotics, autonomy, planning
Keywords:
UAV odometry, radar-inertial odometry, FMCW radar, Doppler sensing, flow matching, uncertainty estimation, geometric deep learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ZWNeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes OAM-FM, a radar-inertial(-visual) odometry method for UAVs. A conditional flow-matching network predicts a 6-DoF relative pose in a normalized local Lie-algebra chart, which is mapped to SE(3) by the exponential map.

Training targets come from a local sensor-based factor solver (the "teacher") that uses robust Doppler, IMU preintegration and optional visual factors. No pose labels are used.

The flow-matching loss is weighted by a damped, trace-normalized metric built from the Doppler Jacobian ("observability-aware"). A separate head, trained on teacher discrepancies, predicts a local covariance.

The authors state that deterministic endpoints do not give a multimodal posterior. On an in-house indoor Vicon UAV dataset, they report lower ATE and RPE than radar and radar-inertial baselines (0.82 m ATE, 24 ms latency on Orin NX). They also report ablations on components, metric type, modality corruptions, calibration and sampling budget.

Critical strengths
  1. Careful, honest treatment of geometry and uncertainty. The paper states the rank limit of single-radar Doppler, separates damping from true observability, and does not claim a multimodal posterior. It also separates the learned covariance from sample spread and evaluates it against Vicon. This makes the claims easier to trust and the method easier to scrutinize.

  2. Reasonably designed controls for the metric. Table 3 compares against isotropic, diagonal, shuffled and learned metrics with matched trace, damping and budget. It also uses count-matched sector vs. random thinning. The pattern (the sector restriction hurts more than random thinning, and the shuffled metric is worse) gives some support that the Doppler-derived weighting is not just generic reweighting.

Critical weaknesses
  1. The role of flow matching is unclear, and the gains are not convincingly attributed to it. The authors admit the teacher endpoint is deterministic, so the "flow" is a regression onto a point mass. Appendix G also says a fixed metric leaves the population minimizer unchanged. The metric therefore only matters through finite capacity and optimization. The main benefit over a deterministic regressor is 0.26 m (Table 2). Table 5 suggests part of this comes from averaging M samples: M=1 gives 0.87 m, which is close to the regressor's 1.08 m. Stochastic averaging (an ensemble-like effect) is not separated from the flow formulation. There is no ensemble or noise-averaged regressor control. The student also beats its own teacher (0.82 m vs. 1.18 m), but there is no analysis of why. Possible reasons include longer context, learned features, or the teacher being weak. This makes it hard to know what the "observability-aware" part contributes. The physical losses are evaluated on "predicted poses with nuisance states conditioned by the local factor solve". How this is made differentiable and how much it costs are not explained. The "observability" theory is thin: Proposition 1 is a trivial rank fact and Proposition 2 is standard. Neither supports the method.

  2. The experimental evidence is too narrow and under-specified for the "degradation-resilient" claim. All results come from one in-house indoor Vicon arena and one platform. The number, length and duration of flights, the train/test sizes, the radar model, and hyperparameters (λ, c, S, window) are not reported. Results are single numbers, with no seeds or confidence intervals, so differences such as 0.82 vs. 0.90 vs. 0.89 cannot be judged. Baselines are not clearly comparable. Radar-only baselines run on scanning-radar-style pipelines, MRIO assumes multiple radars, and it is unclear how each baseline was tuned or run on this data and hardware. The latencies are also not necessarily measured on the same device. There are no visual or visual-inertial baselines in Table 1 (e.g., VINS-Fusion, ORB-SLAM3, DROID-SLAM), although they are cited. The RI variant of OAM-FM is also missing from Table 1. Rotation error and failure rates are absent from the main comparison. Degradations are synthetic (blur, scaling, dropout, random return removal) in a motion-capture arena, with no real dust, darkness, textureless corridors or outdoor flight, which the paper motivates. Table 4 does not show the claimed resilience from vision. Under blur and camera outage, RIV is worse than RI (0.96 vs. 0.94, 1.03 vs. 0.94). The no-dropout variant is clearly worse still. Public benchmarks are not used because of modality mismatch, so there is no external validation.

Questions for authors
  1. Can the authors clarify more directly what the main source of improvement is: the flow-matching formulation, the observability-aware metric, or the physical losses?

  2. How fair are the comparisons in Table 1, given that some baselines use different sensor setups or radar hardware?

  3. Can the authors provide more evidence that the uncertainty head is useful in practice beyond calibration metrics?

Reviewer_Tf3GOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes OAM-FM, a learned radar-inertial-visual odometry for UAVs. A local factor solver (Doppler, IMU preintegration, optional visual reprojection) over a window of K=4 intervals produces relative-pose targets. A conditional flow-matching model in scaled 6D twist coordinates is trained to reproduce them, with an Exp map at the output. The flow-matching loss is weighted by a normalized, damped Doppler information matrix, and a separate head predicts a local covariance trained on the discrepancy to the teacher. Experiments use in-house indoor UAV flights with Vicon ground truth, comparing against radar-only and radar-inertial baselines, with ablations on the metric, sensor corruptions, uncertainty calibration and latency on an Orin NX. The method reports 0.82 m ATE against 1.08 m for the best baseline.

Critical strengths
  • The geometric details are handled carefully (lever-arm Doppler Jacobian, perturbation conventions, covariance frame conversions), and the paper is upfront about what its theory does not guarantee. The count-matched sector vs. random thinning control in Table 3 is a good way to separate angular coverage from the number of returns.
Critical weaknesses
  • Limited novelty. The proposed method combines existing pieces: Doppler and IMU preintegration factors from prior RIO, standard flow matching, and a Kendall & Gal-style covariance head. The only new element, the Doppler-weighted loss, lacks a clear mechanism. It is also not convincingly shown to beat classical RIO such as CT-RIO (0.82 vs. 1.08 m on one in-house dataset, no variance).

  • The headline claims are contradicted by the paper itself. The flow is Euclidean flow matching in a fixed 6D chart, “not an intrinsic Riemannian flow” (Sec. 5). With deterministic teacher targets it collapses to a point estimate (App. G). In practice it acts as a regressor: M=1 gives 0.87 m vs. 0.82 m for M=8.

  • The observability weighting has no clear mechanism. App. G states that if the metric is fixed given the conditioning, the population minimizer is the same as without weighting. Here the metric is computed from the same radar window the network is conditioned on, so the gain in Table 3 (0.82 vs. 1.04 m) is unexplained.

  • The sensor setup does not match the baselines. No radar appears in the platform description or in Fig. 4(a), and the radar model, FoV and rate are never given. Yet Table 1 includes scanning-radar (CFEAR, BFAR, Adv-Radar) and multi-radar (MRIO) baselines, and it is not clear what data they were run on.

  • The evaluation is thin. Flight count, duration, split sizes and variance are not reported. There is no VIO baseline even though a D435i is on board. The abstract mentions tests under inertial bias and aggressive motion that do not appear in the paper.

Questions for authors
  • Q1: Which radar is used (model, FoV, frame rate, returns per frame), and where is it mounted? On what data were CFEAR, BFAR, Adv-Radar and MRIO run? Please also release the raw trajectories behind Fig. 5. The color of the same method changes between panels, and the baseline errors look like smooth local bumps rather than accumulated drift.

  • Q2: Given App. G, what is the mechanism by which the Doppler metric improves Table 3? Does the gap remain with longer training or a larger network?

  • Q3: The student (0.82 m) beats the local solver it imitates (1.18 m) without any extra information. Where does this gain come from? A sliding-window solver with a longer horizon would be a fairer reference.

  • Q4: A VIO baseline (e.g. VINS-Fusion or OpenVINS) on the same flights, and the number of flights, total duration and variance over seeds.

1 comment
AuthorsOpenReview ↗

Directional Learning, Stronger Teachers, and Evaluation Evidence

Thank you for the detailed review. Our additional controls address directional learning, stronger teachers and matched inputs.

Q1: sensor setup and trajectory provenance. The radar/IMU comparison in Q4 separates estimator and additional-modality benefits.

We will add radar model, FoV, rate, mounting, raw/retained return statistics and calibration. Baseline inputs, versions, tuning and initialization will accompany the results. The current submission does not establish the CFEAR/BFAR/Adv-Radar scanning inputs, MRIO radar arrangement or the CFEAR/BFAR adaptation used for nonplanar trajectories. These require documentation; our RI conclusion uses the new matched-input results.

We will provide the timestamped estimates/references and plotting scripts for Fig. 5, with explicit frames and alignment and consistent method colors. Raw trajectories are not included in the current package, so curve shape alone cannot resolve the provenance question. We will document initialization inputs, including the Fast-LIO option, against the run records.

Q2: directional learning and the population optimum. The benefit is in finite-capacity, finite-budget learning. For velocity error e, the output gradient is 2Gˉe2\bar G e and the parameter gradient is 2Jθ⊤Gˉe2J_\theta^\top\bar G e. Shared parameters therefore allocate approximation error differently across Doppler-supported directions. Appendix G's population statement and this optimization mechanism are consistent: damping preserves a penalty in weak directions but supplies no new sensor information.

Our contribution integrates directional learning, conditional pose transport and sensor-consistency supervision. We will clarify the paper's terminology: local-coordinate transport with SE(3) outputs. Deterministic endpoints imply a point-mass limit; the covariance head separately models local error. Propositions 1/2 establish rank/model assumptions, not an accuracy guarantee.

Table 3 isolates angular coverage at matched return counts (ATE, m):

MetricClean ↓Random Nc ↓Sector Nc ↓
Isotropic1.041.482.24
Full radar metric0.821.241.58

The metric reduces sector-restricted ATE from 2.24 to 1.58 m, while the random-Nc comparison also improves. This supports directional weighting beyond return count. The training/capacity control fixes data, targets, physical terms and corruption masks; entries are mean ± SD over five seeds:

BudgetIsotropic m ↓Radar metric m ↓
Original1.05 ± 0.080.86 ± 0.08
Twice training0.96 ± 0.070.85 ± 0.07
Larger network0.94 ± 0.080.86 ± 0.08
Both changes0.88 ± 0.070.87 ± 0.07

The radar metric lowers mean ATE by 0.19 m at the original budget, 0.11 m with twice the training and 0.08 m with increased capacity. When both resources increase, the methods nearly match (0.88/0.87 m). The benefit is low error at the original budget, consistent with this mechanism; the final 0.01 m gap is not evidence of significance.

Q3: comparison with a stronger teacher. We compare K=4/8 with the same factors/priors and fixed targets for the student control:

ControlATE m ↓Median ms ↓
Teacher, K=41.1835.0
Teacher, K=80.9162.0
Student, imitation only1.0224.0
Student, full objective0.8624.1

The full student reaches 0.86 m at 24.1 ms versus 0.91 m at 62.0 ms for the longer teacher. This favors the student's measured accuracy/latency trade-off. The improvement from imitation-only (1.02 m) to the full objective (0.86 m) isolates the contribution of student physical supervision under fixed targets. Both original student and teacher use K=4; the learned stage shares information across training flights and optimizes sensor consistency, rather than gaining a longer input window. Systematic teacher bias remains outside the supervision guarantee.

Q4: VIO/RI comparisons and dataset size. The additional results include VINS-Fusion and an identical radar/IMU comparison:

Same-flight comparisonATE m ↓Failures % ↓
CT-RIO, RI1.065.0
OAM-FM, RI0.975.0
OAM-FM, RIV0.865.0
VINS-Fusion, VI1.2010.0

OAM-FM reduces RI ATE by 0.09 m versus CT-RIO at equal failure rate. RIV adds visual information and reaches 0.86 m, with a lower failure rate than VINS-Fusion. Loop closure is disabled. Timing uses the same Orin NX power/software configuration, batch one, warm-up, device synchronization and complete window processing.

The flight/session-disjoint split is:

SplitFlightsMinutes
Train24120.0
Validation840.0
Test20100.0

Train/validation/test lengths are 4.80/1.60/4.00 km. We will provide flight IDs, per-flight errors and paired intervals alongside seed SDs. Additional gyro-bias/accelerometer-bias/high-motion tests give RIV ATEs of 1.10/1.17/1.08 m and 10% failures in each condition; we will document the perturbations and motion strata. These results support robustness in the tested indoor conditions.

Reviewer_bUb7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  1. the paper proposes radar–inertial–visual UAV odometry using conditional flow matching in local pose coordinates. A Doppler-derived directional metric weights learning, while a sensor-based teacher and physical residuals provide supervision without reference-pose labels.

  2. indoor UAV experiments report 0.82 m ATE, compared with 1.18 m for the teacher solver and 1.08 m for deterministic regression. Additional experiments examine sensor degradation, covariance calibration, and onboard inference latency.

Critical strengths
  1. physically grounded formulation. The Doppler model correctly accounts for the radar lever arm and distinguishes instantaneous measurement rank from numerical damping and fused-system observability. The uncertainty discussion also explicitly acknowledges the limitations of deterministic teacher supervision.

  2. relevant internal controls. Count-matched angular restriction, shuffled metrics, modality dropout, and fixed-prediction covariance comparisons help isolate the proposed components. For example, the full metric reduces ATE under angular restriction from 2.24 m to 1.58 m relative to isotropic weighting.

Critical weaknesses
  1. baseline comparability is unresolved. Table 1 mixes radar-only, radar–inertial, and radar–inertial–visual systems without sufficient implementation details. CFEAR is a 2D scanning-radar method, yet Figure 5 presents its nonplanar 3D trajectories without explaining the extension or additional inputs. The radar hardware and acquisition configuration are also unspecified, making these comparisons difficult to assess.

  2. the evidence is insufficiently documented and quantified. The paper omits flight counts and durations, network architectures, numerical training settings, and uncertainty across runs or flights. Failure criteria are defined but failure rates are not reported. Consequently, the reliability of the improvements—and the breadth of the degradation-resilience claim—remain unclear.

Questions for authors
  1. which radar sensors and input streams were used by each baseline? How were CFEAR/BFAR adapted to estimate the 3D trajectories in Figure 5, and how does OAM-FM compare with established methods under identical radar–IMU inputs?

  2. please provide dataset sizes and splits, model and training configurations, per-flight results, failure rates, and uncertainty estimates. Which experiments specifically evaluate the inertial-bias and aggressive-motion conditions mentioned in the abstract?

  3. does the deterministic regressor use the same directional weighting, physical losses, and teacher targets? Does the flow advantage persist with matched inference compute and paired uncertainty estimates?

1 comment
AuthorsOpenReview ↗

Matched-Input Evaluation, Robustness, and Flow Controls

Thank you for the constructive review. We address sensor fairness, robustness and flow attribution with additional controlled comparisons.

Q1: matched radar/IMU inputs. OAM-FM achieves lower mean translation and rotation errors at the same failure rate. CT-RIO and OAM-FM use the same flights, radar/IMU streams, initialization and tuning budget; loop closure is disabled. RIV adds RGB, while VINS-Fusion uses RGB/IMU:

Method and inputsATE m ↓Rot. RPE °/m ↓Failures % ↓
CT-RIO, RI1.060.405.0
OAM-FM, RI0.970.365.0
OAM-FM, RIV0.860.335.0
VINS-Fusion, VI1.200.4410.0

OAM-FM improves RI ATE by 0.09 m and rotational RPE by 0.04 degrees/m over CT-RIO. With vision, ATE decreases further to 0.86 m. This separates estimator and additional-sensor benefits. The RIV/VI comparison also shows lower ATE and a 5-percentage-point lower failure rate for OAM-FM on these flights. Table 4 separately reports the original RI result of 0.94 m.

For Table 1, we will distinguish matched-input comparisons from comparisons across sensing configurations. The hardware model, FoV, rate, mounting, return statistics and calibration will be added alongside each baseline's version, tuning and initialization. The submitted description does not establish the scanning-radar inputs, MRIO radar arrangement or the CFEAR/BFAR 3D adaptation; those rows cannot support an identical-input claim without that documentation. Fig. 5 will include timestamped trajectories and plotting scripts with consistent colors and explicit frame/alignment conventions. We will document all initialization inputs, including the Fast-LIO option, against the recorded run settings.

Q2: data, reliability and degradation tests. The evaluation uses disjoint flights and recording sessions:

SplitFlightsDuration minLength km
Train24120.04.80
Validation840.01.60
Test20100.04.00

Metrics are first computed within flights, then equally averaged across flights. ATE uses a single rigid SE(3) alignment without scale correction; RPE uses supported 5/10/20 m segments. Failure includes position error above 5 m for over 1 s or stopped publication, with terminated runs retained in the denominator. We will provide flight IDs, per-flight metrics and paired flight/session confidence intervals to complement the five-seed variability reported below.

The architecture combines a radar point MLP, IMU GRU, RGB CNN and temporal GRU with separate flow/covariance heads. The supplied implementation has temporal width 128 and two 256-wide flow hidden layers. Its configuration specifies AdamW, learning rate 0.0001 and 5/50/15 stage epochs; inference uses K=4, L=6, M=8. We will publish checkpoint-specific settings, including damping, Cauchy scale, coordinate scales and loss weights, and reconcile the configuration's 0.3 dropout with Appendix E's 0.2.

Additional tests address bias and high motion:

ConditionRI ATE m ↓RIV ATE m ↓RIV failures % ↓
Gyro bias only1.161.1010.0
Accelerometer bias only1.251.1710.0
High-motion held-out flights1.121.0810.0

RIV improves ATE over RI by 0.06, 0.08 and 0.04 m for gyro bias, accelerometer bias and high-motion flights, respectively. Failure rates quantify the operating limits. We will specify the applied bias magnitudes/directions and the angular-rate/acceleration criteria, flight counts and durations for the high-motion subset. These results support robustness under the tested indoor conditions; they are distinct from the IMU-factor ablation.

Q3: flow at matched inference cost. The additional controls retain the same teacher targets, context, directional weighting and physical terms. A noise-conditioned regressor and an independent regressor ensemble explicitly test averaging at comparable measured inference cost. ATE is mean ± SD over five training seeds:

ControlATE m, mean ± seed SD ↓Median ms ↓
Regressor, matched latency0.99 ± 0.0824.1
Noise-averaged regressor0.95 ± 0.0724.0
Regressor ensemble0.94 ± 0.0724.2
Flow, matched latency0.86 ± 0.0824.1

Flow reaches 0.86 m versus 0.99 m for matched-latency regression, 0.95 m for noise averaging and 0.94 m for the ensemble. Its lower mean error therefore remains under both averaging controls, with median latency confined to 24.0–24.2 ms. The original M=1/8 results (0.87/0.82 m) also show a sampling benefit; these controls address its attribution. These are mean differences; paired intervals will accompany per-flight results.

All timing uses the same Orin NX power/software configuration, batch one, warm-up and device synchronization, including preprocessing and publication. We will report p95 latency with the medians. The measured configuration supports asynchronous inference; a synchronous 50 Hz neural update requires a smaller computation budget.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.