DyRAD renders full range-azimuth-Doppler radar tensors of dynamic driving scenes through a fixed point-spread function, raising radar detection recovery on RADIal from 26.9% to 90.7% of reference-detected objects

DyRAD is described by the authors as the first radar novel-view synthesis method to render Doppler for dynamic driving scenes: the scene consists of static background point reflectors and dynamic point reflectors following rigid object tracks, with reflector velocities derived from those tracks and projected onto the line of sight to render complete RAD tensors. Prior radar novel-view synthesis methods that address dynamic scenes reconstruct only range-azimuth (RA) tensors and omit Doppler, while methods that render Doppler assume static scenes where it arises solely from sensor motion; DyRAD makes Doppler both a rendered output and a constraint on object motion. The paper evaluates on RADIal, Boreas, and a self-built synthetic benchmark; ablations show Doppler supervision reduces object Doppler peak error from 3.2 to 1.1 bins on-path and from 3.4 to 1.3 bins off-path, and raises joint detection F1 from 0.41 to 0.62 on-path and from 0.38 to 0.55 off-path.

DyRAD renders each point reflector through a fixed analytic point-spread function derived from the sensor's signal-processing chain, separating sensor-induced measurement spread from the learned scene structure. Existing methods learn the spatial extent of occupancy or transmittance structure directly from measurements, so a broad return can be explained by a broad scene element rather than the sensor PSF; once absorbed into the scene representation, that spread behaves like static geometry and renders incorrectly from displaced viewpoints, whereas DyRAD uses zero-extent reflectors with a fixed sensor response. Ablations show the fixed analytic PSF more than doubles joint detection recall relative to learned Gaussian extents or learned PSF bandwidths, with off-path F1 reaching 0.55 versus 0.31 for learned extent and 0.29 for learned bandwidths.

This separation also enables zero-shot sensor-configuration transfer, allowing the same reconstructed scene to be rendered under different radar specifications without refitting. Prior radar novel-view synthesis methods cannot reuse the same scene after a sensor configuration change because the sensor response is entangled with scene structure. Using coarse and fine configurations processed from the same RADIal raw ADC recordings, after fitting to coarse measurements and replacing only the PSF and sampling grid, direct rendering reduces Chamfer distance from 1.42 to 1.18 and increases F1 from 0.31 to 0.58 over linear upsampling of the same model's coarse renders.

The paper introduces off-path evaluation protocols, using displaced ground-truth views in a synthetic benchmark and a cycle-consistency protocol adapted to real radar recordings, to test synthesis at displaced viewpoints. Existing radar novel-view synthesis evaluates only at held-out frames along the driven trajectory, where test frames sit between nearly identical training poses so that interpolating directly in measurement space can score as well as a correct forward model, failing to test the displaced-viewpoint synthesis that closed-loop simulation requires. Across 25 scene-offset pairs, DyRAD achieves the highest mean correlation, with object-region RAD correlation of 0.62 versus 0.41 for DyRAD-static and 0.33 for RadarFields; on real data under the Neural LiDAR Fields cycle-consistency protocol, it recovers 90.7% of reference-detected objects on RADIal versus 26.9% for the strongest baseline and 33.1% for the static model.

The work targets sensor simulation that reconstructs dynamic driving scenes from recorded data and synthesizes radar observations at new poses, applying to a Doppler-capable front-facing radar (RADIal) and a Doppler-free spinning radar (Boreas), and relying on dynamic-object bounding-box annotations to initialize tracks. It lets the same reconstructed scene be rendered under different radar processing configurations without refitting, and provides a synthetic benchmark with displaced ground truth for reuse. For practitioners, this means radar observations can be generated beyond the original driven trajectory for closed-loop evaluation and data synthesis, provided the measurements lack elevation and the scene is modeled in the ground plane.

Reflector-to-object assignments rely on object annotations and remain fixed during optimization even as tracks and reflector positions are refined; the authors suggest jointly refining these assignments could help correct initialization errors. Multipath and speckle noise are not modeled, which may affect measurement realism. The current implementation uses a ground-plane model because the evaluated measurements lack elevation, and the authors suggest future work may evaluate elevation-resolving radar and explore transfer across physical sensors. The synthetic benchmark uses an idealized sensor model and does not reproduce real DDMA demultiplexing artifacts, so absolute off-path accuracy still needs testing on more real sensors. In addition, several table values appear as blanks in the loaded text, so some specific numbers cited in the prose cannot be cross-checked against the tables, which limits the detail that can be verified.