Adverse weather and poor illumination remain major challenges for robust ego-trajectory planning in mobile autonomy. 4D radar offers reliable sensing under adverse conditions and direct radial-velocity measurements. However, sensing robustness does not necessarily translate into robust downstream planning, while existing 4D radar benchmarks focus primarily on perception rather than trajectory planning. We present Radar2Plan, a modular benchmark for open-loop ego-trajectory planning using real-world 4D radar data. Radar2Plan connects sensor encoders, scene representations, and planning heads through common interfaces, enabling controlled comparisons between different sensor and planner configurations. Using DSERT-RoLL and MAN TruckScenes, we evaluated seven combinations of camera, 4D radar, and LiDAR in four representative planning baselines and 12 weather and illumination conditions under a unified protocol. To our knowledge, Radar2Plan is the first benchmark dedicated to evaluating real-world 4D radars for autonomous driving planning. Experiments show that 4D radar alone supports competitive ego-trajectory planning and robust performance under adverse conditions. Sensor-configuration comparisons further demonstrate its complementary value to other modalities, while revealing dependencies on sensor combination and planning architecture. The modular design also supports additional datasets, sensing modalities, and planners, providing a flexible foundation for future research on radar-based autonomous driving planning.
Figures & tables
Benchmark
Real sensor logs
Real 4D radar
Ego planning
Long range
Condition eval.
Modality controls
Closed loop
nuScenes [ 1 ]
✓
✗
✗
✗
✗
✗
✗
nuPlan [ 2 ]
✓
✗
✓
✗
✗
✗
✓
Waymax [ 3 ]
✗
✗
✓
✗
✗
✗
✓
Bench2Drive [ 4 ]
✗
✗
✓
✗
✓
✗
✓
K-Radar [ 5 ]
✓
✓
✗
✗
✓
✗
✗
View-of-Delft [ 6 ]
✓
✓
✗
✗
✗
✗
✗
TABLE I: Comparison of representative planning benchmarks and 4D radar datasets.
Fig. 1: Qualitative illustration of how 4D radar sensing robustness benefits downstream planning. (a)–(e) Projected 4D radar returns remain reliable or hardly affected under clear, snow, night, and heavy rain conditions. (f) Ego-frame bird’s-eye-view (BEV) of the heavy rain scene in (e). A SparseDrive-style sparse-query planner [ 12 ] predicts Top-1 trajectories in camera-only, radar-only, LiDAR-only, and camera-radar sensor configurations. The radar-only and camera-radar predictions remain close to the logged trajectory despite degraded visual observations.
Fig. 2: Overview of Radar2Plan . (A) Dataset-specific adapters construct planning samples for seven configurations of camera ( C ), 4D radar ( R ), and LiDAR ( L ). (B)-(D) Modality-specific encoders and planning-specific scene adapters connect these inputs to four representative planning baselines. (E) A common interface returns scored trajectory candidates for evaluation.
Dataset
Setting
Samples
train/val/test
condition groups
DSERT-RoLL
Urban/suburban
10,650
7,052 / 1,305 / 2,293
6
MAN TruckScenes
Truck/highway/terminal
16,403
13,113 / 1,643 / 1,647
6
TABLE II: Dataset statistics under the Radar2Plan planning protocol.
Fig. 3: Qualitative visualization of 4D-radar-based ego-trajectory planning on DSERT-RoLL under heavy rain and MAN TruckScenes under overcast conditions. Insets show the predicted and ground-truth BEV trajectories.
DSERT-RoLL
MAN TruckScenes
Input
GRU
Sparse Query
Vocabulary
Diffusion
GRU
Sparse Query
Vocabulary
Diffusion
C
2.71
2.95
3.20
3.06
5.54
5.78
5.75
7.03
R
2.55
2.45
2.71
2.51
2.10
1.96
2.11
2.18
L
2.71
2.52
2.81
2.53
6.23
6.11
11.81
6.43
C+R
2.57
2.38
3.04
2.86
2.26
2.36
2.25
2.26
C+L
2.76
3.24
3.06
2.91
5.21
4.67
5.70
6.96
TABLE III: Overall Top-1 ADE (m) across matched sensor configurations. Lower values means better results.
Fig. 4: Condition-stratified six-second Top1ADE on DSERT-RoLL and MAN TruckScenes. Rows within each dataset denote sensor configurations, and columns denote dataset-specific weather or illumination conditions.
Fig. 5: Matched 4D radar gain in six-second Top1ADE on DSERT-RoLL and MAN TruckScenes. For each planning baseline, the three comparisons add radar to camera, LiDAR, and camera–LiDAR inputs, respectively. Radar gain is defined as A(f)(X)−A(f)(X+R) according to Eq. ( 6 ), so positive values indicate lower planning error after adding radar.