Learning Where Doppler Velocity Fails: Continuous-Time Reservoir Corrections for Sparse Radar

Radars on moving vehicles measure two things that get tangled together: how far away an object is, and how fast the sensor is moving relative to it. To get clean maps and accurate localization, the system has to estimate its own velocity and subtract that motion's effect from every scan. When the radar sees plenty of stationary structure, this velocity follows almost in closed form from physics. When it does not, the estimate quietly goes wrong. That happens when scans drop out, when only moving cars are visible, inside a featureless tunnel, or when reflections bounce off the ground. We propose a small learned layer that sits on top of a standard physics-based filter. It reads the filter's own stream of surprises, which arrive at irregular times, and outputs a correction and an honest uncertainty for the velocity. The layer's internal dynamics are random and frozen, so the only trained part is a linear readout fit by one regularized least-squares solve.

Stated for a colleague in two sentences: keep the physics for velocity estimation, and add a fixed random continuous-time recurrent system whose linear readout learns the systematic errors the physics makes when radar data becomes sparse. Because only a linear layer is trained, it costs one matrix-vector update per sensor event and can be recalibrated for a new vehicle in seconds. The open question is whether this beats a well-tuned filter in exactly the sparse regimes where filters fail, while doing no harm when data is plentiful.

Core Claim

The literature already establishes that velocity-based Doppler compensation matters. In reverse-path localization on Boreas, adding Doppler correction to motion de-skewing cut mean translation error from 2.3 m to 0.43 m (Do We Need to Compensate for Motion Distortion and Doppler Effects in Spinning Radar Navigation?). The weak point is the velocity estimate itself in degraded conditions. The learned velocity regressor of Doppler-aware Odometry from FMCW Scanning Radar systematically underestimates speed and covers only forward motion. The Doppler loss in RadarMOSEVE breaks down when no static background is visible. Continuous-time Gaussian-process estimators (STEAM-RIO, Go-RIO) handle asynchronous timing well. However, they bridge gaps with a generic constant-velocity-style prior and require iterative solvers.

Our claim has three parts. First, the errors of a physics Doppler-plus-IMU estimator under sparsity are structured: they depend on gap length, the fraction of static points, recent maneuvers and IMU bias history. Second, a random continuous-time reservoir driven by the irregularly timed stream of filter innovations can expose that structure to a linear readout. Third, the corrected velocity and inflated covariance, fed back into Doppler de-skewing, reduce downstream drift and localization error under controlled radar dropout, at a fraction of the cost of a trained recurrent model. The reservoir does not replace the physics estimator. A reviewer would rightly point out that full-replacement is weaker, since a static-point least-squares fit already solves the well-conditioned case in closed form.

Research Program

The base estimator is a conventional extended Kalman filter (EKF), a filter that propagates a state with IMU readings and corrects it with measurements. It takes per-scan ego-velocity from a least-squares fit to the radial Doppler of static points, with RANSAC rejecting outliers. For spinning FMCW radar without a native Doppler channel, the velocity measurement comes from MC-RANSAC-style motion estimation. Every radar update produces an innovation, meaning the difference between the predicted and measured velocity, at an irregular time. Alongside it we record the inlier fraction, the time since the last valid scan, and the high-rate IMU signal.

These channels form a control path $X_t$ that drives a Random Fourier CDE in the style of Random Controlled Differential Equations. That paper evolves a large hidden state through random, frozen dynamics and fits only a ridge-regression readout. It reports average accuracy 0.741 on 16 UEA classification datasets and stable behavior up to 40% missing data. It was never applied to physical sensing, and we reuse its machinery here. Concretely,

$$dh_t = \sum_{i} \left(A_i h_t + b_i\right) dX^i_t, \qquad \Delta \hat v_t = W h_t,$$

Here $h_t$ is the reservoir state, $X^i_t$ is the $i$-th input channel after a random Fourier lift, $A_i$ and $b_i$ are fixed random matrices and biases, and $W$ is the only trained quantity. $W$ maps the state to a 3D velocity correction $\Delta \hat v_t$ against GNSS/INS ground truth by ridge regression. A second ridge readout predicts the log squared residual, which inflates the EKF measurement covariance when the reservoir expects the Doppler fit to be unreliable. This is what makes the compensation adaptive. Because the state evolves only between events, irregular radar timing needs no imputation or grid binning. For the rough, high-rate IMU channel we test the log-ODE windowing of the Random Rough DE variant from the same paper.

The program has three parts. The first builds the corrected estimator and plugs its velocity into existing Doppler de-skewing pipelines. The second analyzes what the readout learns, by regressing its corrections against gap length, inlier fraction and speed; this checks whether it captures the known underestimation bias rather than dataset idiosyncrasies. The third tests cheap transfer: refitting only $W$ on a few minutes of data from a new platform, for example moving from Boreas's spinning radar to a 4D imaging radar, while keeping the reservoir untouched. This last test draws on the continual-learning argument for fixed reservoirs in Sustainable Deep Learning for Time-series: Reservoir Computing and the fixed-pole stability results of When Learning Hurts. Neither of those works addresses irregular time gaps or Doppler physics.

The work also takes on the noise problems described in ORORA and DC-Loc. Both show that radar measurement uncertainty is strongly direction-dependent and that outliers can be extreme. Existing continuous reservoirs treat inputs as clean and treat all input directions the same way. We sidestep this by feeding innovations that have already been whitened by the filter's covariance, rather than raw points. We report whether that suffices.

Baselines

The naive reference is the physics estimator with no Doppler correction, which uses scan matching on uncompensated scans. The second reference is physics-only Doppler compensation, with velocity from per-scan static-point least squares or MC-RANSAC and nothing learned. On Boreas reverse-path localization, the latter has a reported mean translation error of 0.43 m (Burnett et al., 2021), and that is the number our corrected velocity must improve on in the clean setting.

The strongest practical competitor is a tuned EKF that fuses IMU and Doppler velocity, with innovation-based noise adaptation. This is the cheapest serious alternative for the same job. No reported result was found on Lacuna for EKF-RIO's absolute error, and Go-RIO reports beating it without baseline values. We will therefore run it ourselves, tuning its noise parameters by grid search with the same number of trials the reservoir gets for its scale and regularization parameters.

The strongest accuracy baseline is continuous-time Gaussian-process radar-inertial odometry. STEAM-RIO reports a 43% reduction in translational drift on Boreas, but the file gives no absolute drift value. Go-RIO reports 0.72% vertical drift on RURAL_A2. We reimplement this baseline on the same sequences and dropout masks, and report accuracy and per-update latency together. On RDD the reference is the fused scan-matching plus DAVE-NN system at 4.59% translation error (Rennie et al., 2023). On View-of-Delft ego-velocity, the reference is RadarMOSEVE's 94.3% precision at 0.5 m/s (RadarMOSEVE).

Three ablations isolate the key ingredients. The first replaces random dynamics with a trained GRU or Neural CDE of equal state size on identical inputs. It gets the same training data and a matched wall-clock budget. The second replaces continuous time with a discrete echo state network that receives the time gap as an extra input, in the style of Temporal Belief Memory. The third removes the covariance readout, keeping the correction only.

Beating a baseline means a statistically reliable reduction in velocity error under at least moderate dropout, with no degradation of more than a few percent on clean data. It also means lower downstream drift or localization error than the tuned EKF, and per-update cost below the Gaussian-process estimator. Matching the trained GRU while training in seconds counts as a success for the cost claim but not the accuracy claim.

Why not just…?

Why not just use a well-tuned adaptive EKF? This is the real threat. On clean data we expect it to match us, and the proposal is built so that this outcome is acceptable. The bet is that its errors under sparsity are history-dependent: bias accumulated during a tunnel, or underestimation correlated with low inlier fraction. A memoryless noise-adaptation rule cannot represent that kind of dependence. The burst-dropout and static-masking experiments settle the question directly. If the adaptive EKF closes the gap there, the reservoir layer is unnecessary, and we will report that.

Why not just use the continuous-time Gaussian-process estimator? It already handles asynchrony, and Burnett et al., 2024 shows a large drift reduction. However, its prior is a generic white-noise-on-acceleration model that does not learn vehicle- or sensor-specific error patterns. It also needs an iterative solver; its lidar variant takes about 74 ms per frame. We compare both accuracy and latency. We also test adding our correction on top of it, since the two are complementary rather than exclusive.

Why not just train a recurrent network or Neural CDE end to end? It might be more accurate. Random Controlled Differential Equations shows random reservoirs roughly matching trained kernel methods on benchmarks, but that is not evidence about radar. The trained-dynamics ablation measures the accuracy gap directly. The transfer experiment then measures whether refitting only a readout recovers performance on a new platform faster than fine-tuning.

Why not just impute the missing scans and run an LSTM? Forward filling erases the information carried by when data is missing (He et al.). Here, gap length is precisely the signal that predicts error. We include a forward-filled LSTM on the same inputs as a check.

Proposed Experiments

We use Boreas for spinning-radar odometry and reverse-path localization, RDD for feature-poor tunnels, NTU4DRadLM and MSC-RAD4R for 4D radar with IMU, and View-of-Delft for ego-velocity. On each, we impose controlled sparsity: random scan dropout from 0% to 60%, burst dropouts of 1 to 10 seconds, and masking of static points down to 5% to imitate scenes dominated by moving objects. Metrics are 3D velocity root-mean-square error, KITTI-style translational and rotational drift, reverse-path mean translation error, calibration of the predicted variance (the fraction of true errors that fall within predicted intervals), and per-update latency on an embedded CPU.

Four checks decide the project. At 0% dropout, the corrected estimator must not worsen the physics-only 0.43 m Boreas localization. Under burst dropout and static masking, it must beat the tuned adaptive EKF on velocity error and on downstream drift. The continuous-time reservoir must beat the discrete echo state network with time-gap inputs, which isolates the value of continuous time. Readout-only refitting on a new radar type must recover most of the in-domain gain from minutes of data. If the second check fails, the core claim is false.