MulDar

Unleashing the Potential of Distributed COTS mmWave Radar by Exploiting Cross-Device Channels

MobiSys 2026

Xinghua Sun, Qiancheng Li, Akshay Gadre

University of Washington

Paper Video Code 🤗 Dataset DOI

MulDar develops a synchronous multi-static radar system that addresses specular reflection and ghost artifacts, making previously invisible objects detectable in mmWave imaging.

MulDar transforms a fleet of distributed commodity radars into a single coherent sensing array, surpassing the capabilities of any individual sensor.

More information. With N radars, every transmitter can talk to every receiver, so the number of sensing channels grows quadratically rather than linearly.
More perspectives. Spatially separated radars view the scene from different angles, so a surface that deflects one radar's signal still reflects toward another.
No extra infrastructure. Each radar runs on its own internal clock; no synchronization cables or shared oscillators are needed.

The result is a far more complete and accurate picture of the scene. MulDar reliably captures mirror-like surfaces and objects hidden behind clutter, the very cases where a single radar fails.

Key Contributions

Distributed mmWave synchronization. MulDar engineers a direct over-the-air path between commodity TI AWR mmWave radars for coarse hardware-trigger synchronization and per-chirp phase/frequency/amplitude offset correction, enabling cross-device channel capture on COTS hardware with no hardware modifications.
Multi-static bi-static imaging. A backprojection-based bi-static SAR imaging algorithm reconstructs the 2D scene from cross-radar channel measurements. Correlated channel pairs are coherently fused for SNR and resolution gain; uncorrelated pairs are non-coherently fused to add new viewpoints.
Specular reflection recovery & ghost removal. By combining mono-static and bi-static measurements, MulDar detects objects invisible to any single radar and removes ghost multi-bounce artifacts — achieving a 66.79% reduction in Chamfer distance vs. mono-static baselines across planar surfaces, deformed reflectors, and real-world scenes.

Motivation

Mono-static mmWave channels are fundamentally limited by: sparsity and noise. Most surfaces act as specular mirrors at mmWave wavelengths, so a co-located TX and RX only sees energy from the narrow set of surfaces whose normal points back at the radar. Everything else reflects energy harmlessly away, leaving a sparse return in the received signal.

Fundamental limitations of mono-static mmWave channels: sparsity and noise

System Overview

MulDar system overview

Distributed radars transmit in a round-robin schedule, the direct path between radars is used for phase synchronization, and a backprojection algorithm fuses mono-static and bi-static measurements for full scene reconstruction.

MulDar synchronizes bi-static with direct-path referencing

The bi-static channels is not perfect which introduces random frequency and phase offsets. MulDar corrects these offsets using the direct-path peak as a reference, enabling coherency across radars for multi-static imaging.

Channel fusion strategy: correlated channels are coherently combined for higher resolution and SNR, uncorrelated channels are non-coherently fused to add new viewpoint information.

Results

We evaluate MulDar against two baselines: Mono, which fuses only mono-static reflections from each radar, and Multi, which fuses only the cross-radar bi-static channels. Across planar surfaces of different materials, deformed metal reflectors, and cluttered real-world scenes, MulDar consistently recovers geometry that both baselines miss. Mono-static radars see specular surfaces only as point reflections (or not at all), while bi-static-only fusion misses objects whose dominant return is mono-static. MulDar's joint fusion captures both, producing complete and ghost-free reconstructions.

Qualitative comparison across planar surfaces, deformed reflectors, and real-world scenes. MulDar (right) reconstructs objects that are invisible or ghost-distorted in the mono-static (left) and multi-static-only (center) baselines.

MulDar has better coverage to the objects. We measure the two-way Chamfer distance between the reconstructed point cloud and the ground truth across planar surfaces of five materials: metal, plastic, wood, fabric, and drywall. They were rotated through a wide range of incidence angles. MulDar reduces the average Chamfer distance by 72.62% on metal, 81.03% on plastic, 69.47% on wood, 68.47% on fabric, and 42.38% on drywall, for an overall 66.79% improvement over the mono-static baseline.

Average Chamfer distance across planar surfaces of five materials. MulDar achieves a 66.79% reduction over the mono-static baseline.

MulDar in both indoor and outdoor scenarios.

Indoor evaluation: MulDar images corridor corners, open areas, and pillars that mono-static radars miss entirely.

Automotive evaluation: MulDar detects the specular surfaces of a neighboring car that are invisible to conventional mono-static mmWave radars.

BibTeX

If you find this work useful, please consider citing our paper!

@inproceedings{sun2026muldar,
  author    = {Sun, Xinghua and Li, Qiancheng and Gadre, Akshay},
  title     = {MulDar: Unleashing the Potential of Distributed COTS mmWave Radar by Exploiting Cross-Device Channels},
  booktitle = {Proceedings of the 24th Annual International Conference on Mobile Systems, Applications and Services (MobiSys '26)},
  year      = {2026},
  doi       = {10.1145/3745756.3809206},
  publisher = {ACM},
  address   = {New York, NY, USA},
  month     = {June},
}