arXiv 2511.14019  ·  November 2025

RISE Single Static Radar-based
Indoor Scene Understanding

One static mmWave radar. Full room layouts and object boxes — including the parts the radar can never see directly.

Signal Kinetics Group  ·  MIT Media Lab

Paper PDF Code Dataset
RISE overview: conventional mmWave detection sees only a limited region, while RISE uses multipath to reconstruct the full layout and detect occluded objects.
Conventional single-radar sensing (left) sees only the directly visible slice of a room. RISE (right) treats wall reflections as extra viewpoints, recovering geometry and objects in regions the direct path never reaches.
16cm
Layout Chamfer distance
60% lower than prior state of the art
58%
Object detection IoU
first object-level mmWave detection
1radar
Static, single sensor
no motion, no arrays of devices
50K
Annotated radar frames
100 real indoor trajectories
Abstract

Indoor scene understanding is fundamental to numerous applications, yet existing approaches rely on optical sensors that struggle with occlusion and raise privacy concerns. Millimeter-wave radar is a compelling privacy-preserving alternative — it penetrates obstacles and works regardless of lighting. But its low spatial resolution and the complex multipath propagation of indoor environments make accurate scene understanding hard.

We introduce RISE, a system that jointly performs indoor layout reconstruction and object detection from a single static mmWave radar. Our key insight is to leverage multipath reflections as geometric information rather than suppress them as noise. RISE employs Bi-Angular Multipath Enhancement (BAME), which explicitly models both Angle-of-Arrival and Angle-of-Departure to recover secondary reflections and reveal hidden structure, paired with a Sim2Real Hierarchical Diffusion (SRHD) framework that turns fragmented observations into complete scene representations.

We also introduce the first large-scale radar dataset for this task — 50,000 frames across 100 real indoor trajectories. RISE reduces Chamfer distance by 60% over prior work, reaching 16 cm accuracy, and delivers the first mmWave-based object detection at 58% IoU.

Key insight

Multipath is not noise. It is a second viewpoint.

A signal that bounces off a wall before reaching an object arrives late and from the wrong direction, producing a “ghost” that classical pipelines discard. Every ghost, however, encodes the wall that created it — and a view of the scene from that wall.

Multipath-induced ghost detections and the corresponding signal strength map.
Multipath-induced ghosts. (a) Reflections off a wall create phantom targets. (b) The signal-strength map shows the spurious peaks. RISE inverts this effect to recover the true geometry.
Conventional

Ghosts are rejected as clutter. The scene collapses to whatever the direct path can reach — typically one wall and any object in clear line of sight.

RISE

Each ghost is assigned to the surface that produced it, then folded back to its true position. Occluded corners and objects behind furniture re-enter the reconstruction.

Method

From raw chirps to a complete scene

Two components in sequence: a signal-processing front end that recovers geometry from multipath, and a hierarchical diffusion model that completes what the radar could not measure.

RISE pipeline: BAME signal processing followed by Sim2Real Hierarchical Diffusion.
System pipeline. Raw mmWave captures pass through BAME for geometric recovery, then through SRHD for scene completion.
  1. 01Range FFT
  2. 02AoA / AoD FFT
  3. 033D CFAR
  4. 04Ghost reintegration
  5. 05Multipath inversion
  6. 06Temporal accumulation
  7. 07GMM + RANSAC
  8. 08Layout diffusion
  9. 09Object diffusion
Component 01

Bi-Angular Multipath Enhancement

Standard radar measures only the Angle-of-Arrival. BAME additionally resolves the Angle-of-Departure — the direction the signal left the array before it bounced. Jointly reasoning over both angles identifies exactly which surface a return reflected off, turning that surface into a virtual mirror onto occluded space.

  • Joint AoA + AoD estimation per return
  • Closed-form inversion of ghosts to true positions
  • Recovers structure invisible to direct-path sensing
BAME recovers ghost targets missed by conventional AoA-only processing.
BAME recovers returns that AoA-only processing cannot separate.
Component 02

Sim2Real Hierarchical Diffusion

Even after multipath recovery, radar observations stay sparse. SRHD is a diffusion model that denoises accumulated radar evidence into a dense scene: room layout first, then object boxes conditioned on it. A physics-based simulator supplies the training data, and domain adaptation closes the gap to real captures.

  • Hierarchical two-stage generation: layout, then objects
  • Simulator-driven training, no large real annotation burden
  • Joint output: floorplan geometry + object bounding boxes
Two-stage architecture of the Sim2Real Hierarchical Diffusion model.
SRHD conditions object generation on the layout it has already produced.

Multipath inversion

Given a ghost detection and the bi-angular measurement that produced it, RISE solves for the reflecting plane and mirrors the detection back across it. The inversion is closed-form: no learned component, no iterative search — which is what keeps it reliable enough to feed the diffusion stage.

Geometric relationships used in multipath inversion.
The geometry relating a ghost, the reflector, and the true target.
Results

Evaluated on 100 real indoor trajectories

RISE improves over the prior mmWave layout baseline on every trajectory, and adds object-level detection that no previous single-radar system provides.

Wall reconstruction error of RISE versus the prior mmWave layout baseline across 100 trajectories.
Layout reconstruction across 100 trajectories. RISE versus the prior mmWave layout baseline, showing consistent gains rather than a few favourable scenes.
Qualitative comparison: RGB reference, prior baseline, and RISE reconstructions.
Qualitative comparison. RGB reference (left), prior baseline (middle), RISE (right). RISE closes walls that the baseline leaves fragmented and places objects the baseline misses entirely.
Performance of RISE across varying observation trajectory lengths.
Varying trajectory length. Accuracy degrades gracefully as the observation window shortens, which matters for real deployments with limited dwell time.
Dataset

The first large-scale benchmark for this task

Paired radar captures and ground-truth scene annotations across offices, living rooms and corridors — plus the physics-based simulator used to pretrain SRHD.

50,000
radar frames
100
real indoor trajectories
2
annotation types: layout + object boxes
Simulator ground truth and the augmented training data derived from it.
Simulator and augmentation. Top: ground truth from the physics-based simulator. Bottom: augmented samples used to bridge the sim-to-real gap during SRHD training.

Code, trained models and dataset release are hosted on GitHub.

Get the code Read the paper
Why mmWave

What radar buys you

Trade-offs against the optical sensors normally used for indoor scene understanding.

Capability RGB camera LiDAR mmWave radar (RISE)
Works in the darkNoYesYes
Privacy preservingNoNoYes
Sees through obstaclesNoNoYes
Low hardware costYesNoYes
Layout reconstructionPartialYesYes
Object detectionYesYesYes — first for radar
BibTeX
@article{zhou2025rise,
  title   = {RISE: Single Static Radar-based Indoor Scene Understanding},
  author  = {Zhou, Kaichen and Dodds, Laura and Afzal, Sayed Saad and Adib, Fadel},
  journal = {arXiv preprint arXiv:2511.14019},
  year    = {2025}
}