Robust 3D environmental perception is critical for applications such as autonomous driving and robot navigation. However, optical sensors such as cameras and LiDAR often fail under adverse conditions, including smoke, fog, and non-ideal lighting. Although specialized radar systems can operate in these environments, their reliance on bespoke hardware and licensed spectrum limits scalability and cost-effectiveness. This paper introduces Rascene, an integrated sensing and communication (ISAC) framework that leverages ubiquitous mmWave OFDM communication signals for 3D scene imaging. To overcome the sparse and multipath-ambiguous nature of individual radio frames, Rascene performs multi-frame, spatially adaptive fusion with confidence-weighted forward projection, enabling the recovery of geometric consensus across arbitrary poses. Experimental results demonstrate that our method reconstructs 3D scenes with high precision, offering a new pathway toward low-cost, scalable, and robust 3D perception.
Figures & tables
Tech.
Medium
Waveform
O.P.
O.R.
S.L.E.
Hardware
P.C.
Cost
Scalable
Camera
Light
—
No
Poor
—
Dedicated
Low
Low
High
LiDAR
Laser
Pulsed
No
Poor
—
Dedicated
High
High
Low
Radar
Radio
FMCW, etc.
Yes
Good
No
Dedicated
Med
Med
Low
Rascene
Radio
OFDM
Yes
Good
Yes
Reused
Low
∼ Zero
High
Table 1: Comparison of Rascene and conventional modalities. [Keys: O.P.= Obstacle Penetration , O.R.= Occlusion Resilience , S.L.E.= Spectrum License Exempt , P.C.= Power Consumption .]
Figure 2: Illustration of monostatic sensing in a mmWave communication system. The mmWave device (left) simultaneously transmits and receives OFDM communication signal for sensing.
Figure 3: Illustration of angular estimation on a mmWave device.
Figure 4: Examples of generated radio point clouds, where S>a={s∈S∣s>a} . Threshold a is set to 0.1 and 0.2, and the 3D point clouds are projected onto 2D images for for ease of visualization.
Figure 5: Overview of the multi-frame 3D RF imaging network. Given multiple radio frames and poses, a shared encoder predicts per-frame feature maps and confidence logits. We then warp all features to a reference frame and fuse them into a unified representation. A coarse-to-fine 3D decoder with voxel and depth heads outputs the reconstructed voxel grid and depth map.
Figure 6: Rascene data collection platform.
Metric
A
B
C
D
E
F
G
H
Avg.
Depth
AbsRel (%)
10.1
4.9
4.5
3.3
17.3
19.4
7.4
8.7
9.4
MAE (cm)
22.9
11.5
10.2
7.3
36.2
37.4
16.8
19.9
20.2
RMSE (cm)
43.2
21.2
22.2
17.7
63.3
66.9
34.4
38.1
38.0
Voxel
CD (cm)
20.4
15.4
13.0
9.1
37.0
26.5
16.8
16.9
19.7
CD Diag (%)
2.5
1.7
1.5
1.1
3.7
3.3
2.0
2.1
2.3
Table 2: Quantitative evaluation results of Rascene’s cross-scene generalization performance. The model is trained on 12 scenarios and evaluated on 8 distinct, unseen test scenarios (A-H). We report metrics for both depth estimation (AbsRel, MAE, RMSE) and 3D voxel reconstruction (CD, CD Diag ). Lower values are better for all metrics.
Figure 7: Qualitative results of our Rascene system. For each row, we show the ground truth depth map and voxel grid derived from the LiDAR sensor (Cols 1-2), and our corresponding predictions generated using mmWave communication signal (Cols 3-4).
Methods
# of frame
AbsRel
MAE
CD
CD Diag
PanoRadar [ 27 ]
1
14.7%
34.1
32.2
3.8%
CartoRadar [ 28 ]
5
—
—
26.8
3.1%
Rascene (Ours)
1
14.1%
32.9
31.6
3.6%
Rascene (Ours)
5
9.4%
20.2
19.7
2.3%
Table 3: Within-dataset comparison with baselines.
Figure 8: Qualitative comparison of single-frame and 5-frame predictions. Multi-frame fusion reduces missed detections and suppresses hallucinated structures, yielding more complete and geometrically consistent reconstructions closer to the LiDAR ground truth.
Frame #
Depth
Voxel
AbsRel
MAE
RMSE
CD
CD Diag
1
14.1%
32.9
56.2
31.6
3.6%
2
11.1%
24.6
44.2
26.0
3.0%
3
9.8%
21.8
40.4
21.9
2.5%
5
9.4%
20.2
38.0
19.7
2.3%
Table 4: Impact of the number of fused radio frames ( N in Eq. ( 10 )).
Error Range
Depth
Voxel
AbsRel
MAE
RMSE
CD
CD Diag
Rot.
0∘∼5∘
11.7%
25.8
46.8
23.1
2.6%
5∘∼10∘
18.4%
40.0
66.8
31.2
3.6%
10∘∼15∘
18.8%
42.0
67.7
34.3
3.9%
Trans.
0∼5 cm
9.4%
20.2
38.1
19.7
2.3%
5∼10 cm
9.4%
20.4
38.2
19.8
2.3%
Table 5: Impact of pose inaccuracies (rotation and translation perturbations added at test time).
Figure 9: Cumulative distribution functions illustrating absolute and relative depth estimation errors.
Figure 10: Evaluation of depth estimation accuracy across the azimuth and altitude dimensions of the sensor field of view.
Figure 11: Depth error as a function of range for single-frame and multi-frame inference.
Occlusion
Depth
Voxel
AbsRel
MAE
RMSE
CD
CD Diag
No Occ.
5.6%
13.7
24.8
18.4
2.0%
Paper Sheet
6.1%
14.4
25.2
19.1
2.1%
Styrofoam
6.4%
15.7
27.4
21.2
2.3%
Table 6: Evaluation of Rascene’s robustness to common occluders.
Figure 12: Representative examples of different sensors’ occlusion resilience in a corridor scene at different locations.
Figure 13: Our prototyped monostatic ISAC device.
Figure 14: Illustration of video streaming data communication during sensing data collection.
Hardware parameters
Sampling rate
1.2288 GSPS
Number of Tx antennas
16
Number of Rx antennas
16
Center frequency
60 GHz
Transmission power
20 dBm
Communication parameters
Table 7: The parameters of our monostatic ISAC device.
Figure 15: Sample trajectory segments (shown in red) from different scenes, visualized on the ground truth LiDAR point clouds.
Figure 16: Example snapshots from the 20 distinct indoor environments included in our dataset.