Radar2Plan: Benchmarking 4D Radar for End-to-End Open-Loop Ego-Trajectory Planning

Ling Yao; Yichun Xiao; Jin Jin; Yihan Zhang; Fangqiang Ding

Radar2Plan · Research overview

The motivation / sensing

When visibility changes, what can radar still see?

Projected 4D radar returns remain reliable or hardly affected under clear, snow, night, and heavy rain conditions.Does the sensing robustness of 4D radar benefit downstream planning?

Illustrative example DSERT-RoLL · Heavy Rain

Official annotations · Camera and 4D radar sensing. Synchronized camera frames and projected radar returns; boxes are official 2D ground truth, not detector predictions. No planning output is shown.

JavaScript is off. View the clips directly: Heavy Rain ↗ Heavy Snow ↗ Fog ↗ Light Snow ↗ Night ↗ Clear ↗

Robust sensing is the starting point. What about planning?

The research question

Robust sensing.
But robust planning?

Existing 4D radar benchmarks focus primarily on perception rather than trajectory planning.

Despite these advantages, existing 4D radar studies have focused primarily on perception tasks such as object detection, tracking, and occupancy prediction, leaving its value for downstream planning insufficiently evaluated.

We therefore present Radar2Plan, to our knowledge, the first benchmark for open-loop ego-trajectory planning using real-world 4D radar data.

Radar2Plan adopts a modular design that separates dataset adaptation, sensor input encoding, planning-specific scene representation, and the planning method through standardized interfaces.

01 / Standalone

Can a planner generate competitive ego trajectories using 4D radar as its only environmental sensing modality?

02 / Robustness

Under adverse conditions, does radar-based planning achieve lower planning error than camera or LiDAR-based planning?

03 / Complementarity

Does adding 4D radar improve planning when combined with camera, LiDAR, or camera-LiDAR sensor inputs?

Benchmark design

A modular benchmark with standardized interfaces.

This design allows individual components to be exchanged independently while maintaining a consistent evaluation protocol across sensor configurations and planning methods.

Camera, radar and LiDAR pass through dataset adapters, modality encoders, planner-specific scene representations and four planning heads, producing scored six-second trajectories. — Fig. 2 Radar2Plan separates dataset adaptation, modality-specific sensor encoding, planner-specific scene representation, controlled planning baselines, and unified trajectory output and evaluation.

01 / Unified task

Consistent task definitions

Radar2Plan converts DSERT-RoLL and MAN TruckScenes into a unified planning benchmark with consistent task definitions, data splits, and evaluation metrics.

02 / Controlled comparison

Seven sensor configurations

Radar2Plan evaluates seven combinations of camera, 4D radar, and LiDAR in four representative planning baselines.

03 / Unified evaluation

Top-1 ADE at six seconds

Top1ADE@6s averages the displacement error of the Top-1 trajectory over all waypoints up to the six-second horizon.

Datasets, splits & experimental controls

DSERT-RoLL

Urban & suburban driving

Planning samples: 10,650
Train / val / test: 7,052 / 1,305 / 2,293
Conditions: Clear, fog, light rain, heavy rain, light snow, heavy snow

MAN TruckScenes

Highway & terminal trucking

Planning samples: 16,403
Train / val / test: 13,113 / 1,643 / 1,647
Conditions: Clear, overcast, rain, illumination, twilight, dark

We train every configuration for 80 epochs using AdamW. We set the learning rate to 10⁻⁴, the weight decay to 10⁻⁴, and the batch size to 16. All runs use seed 42 without a learning-rate scheduler.

We use Top1ADE@6s as the primary evaluation metric. Unless otherwise specified, Top1ADE refers to Top1ADE@6s.

Perspective 01 / Standalone

Standalone Planning Capability.

Radar-only input achieves the lowest single-modality Top1ADE for every planning baseline on both datasets.

Illustrative example + aggregate charts DSERT-RoLL · Clear

4D Radar only · charts: C / R / L Radar-only input achieves the lowest single-modality Top1ADE for every planning baseline on both datasets. Left: illustrative predictions. Right: full-validation aggregate results, not errors for the displayed sample.

JavaScript is off. View the clips directly: DSERT · Clear ↗ MAN · Overcast / Vocabulary ↗ DSERT · Light Rain ↗ MAN · Overcast / Diffusion ↗

Perspective 02 / Adverse conditions

Planning Robustness under Adverse Conditions.

On DSERT-RoLL, radar-only input outperforms camera-only input in 19 of the 20 adverse-weather baseline-condition pairs; adding radar to camera also improves 19 of the 20 pairs.

Planner

DSERT-RoLL · GRU Top-1 ADE@6s (m) · Lower is better
Sensor input	Clear	Light Rain	Heavy Rain	Fog	Light Snow	Heavy Snow
C	2.22	3.01	2.93	5.05	2.25	7.45
R	2.37	2.19	2.62	2.89	1.77	1.10
L	2.57	2.35	3.05	3.02	1.79	6.50
C+R	2.43	2.81	3.22	4.64	1.45	2.08
C+L	2.64	2.72	3.10	5.01	2.28	2.13
R+L	2.43	2.53	2.90	3.35	3.04	1.00
C+R+L	2.46	2.32	3.37	3.63	2.18	0.93

DSERT-RoLL · Sparse Query Top-1 ADE@6s (m) · Lower is better
Sensor input	Clear	Light Rain	Heavy Rain	Fog	Light Snow	Heavy Snow
C	2.65	2.70	3.41	7.66	2.85	9.41
R	2.28	1.90	3.21	2.94	1.49	1.75
L	2.62	1.97	3.05	2.84	2.53	1.81
C+R	2.05	2.13	3.30	3.77	1.49	5.22
C+L	3.18	3.44	3.77	4.73	1.85	2.88
R+L	3.47	2.26	4.16	4.47	2.32	1.75
C+R+L	3.01	2.63	3.48	3.41	2.45	4.89

DSERT-RoLL · Vocabulary Top-1 ADE@6s (m) · Lower is better
Sensor input	Clear	Light Rain	Heavy Rain	Fog	Light Snow	Heavy Snow
C	4.04	4.76	4.20	7.94	2.39	17.63
R	2.73	3.00	2.95	3.30	2.29	1.91
L	2.93	2.76	3.49	3.61	1.42	1.47
C+R	2.96	3.02	3.26	5.19	1.68	3.11
C+L	2.98	2.99	3.66	3.49	2.74	3.28
R+L	2.53	2.41	3.44	3.33	1.77	1.59
C+R+L	3.42	3.55	3.51	4.96	3.06	0.79

DSERT-RoLL · Diffusion Top-1 ADE@6s (m) · Lower is better
Sensor input	Clear	Light Rain	Heavy Rain	Fog	Light Snow	Heavy Snow
C	3.67	2.91	4.51	4.37	3.29	10.30
R	2.78	2.94	3.06	3.94	2.22	1.50
L	2.60	2.52	2.97	3.86	1.67	0.83
C+R	2.92	2.48	4.14	3.21	2.48	1.39
C+L	3.14	2.72	2.93	4.19	1.37	1.06
R+L	2.50	2.71	2.53	3.21	1.41	0.90
C+R+L	2.60	2.34	3.00	3.87	1.69	2.36

MAN TruckScenes · GRU Top-1 ADE@6s (m) · Lower is better
Sensor input	Clear	Overcast	Rain	Illumination	Twilight	Dark
C	3.62	7.52	3.12	5.75	2.61	4.95
R	1.52	2.64	1.96	2.16	1.69	1.76
L	4.44	8.14	5.81	6.10	5.07	7.62
C+R	1.98	2.50	2.60	2.14	1.65	3.30
C+L	3.81	7.26	2.31	5.20	2.33	6.14
R+L	1.14	2.40	1.62	1.88	0.79	1.09
C+R+L	2.32	2.97	1.88	2.54	2.22	2.86

MAN TruckScenes · Sparse Query Top-1 ADE@6s (m) · Lower is better
Sensor input	Clear	Overcast	Rain	Illumination	Twilight	Dark
C	4.11	7.45	4.30	5.95	3.99	5.13
R	1.44	2.45	1.96	2.07	1.15	1.41
L	4.57	7.50	7.13	6.31	2.52	5.75
C+R	1.90	2.73	3.32	2.39	1.63	2.38
C+L	3.87	5.90	2.72	4.55	2.58	6.12
R+L	1.48	2.43	1.83	2.07	1.34	1.44
C+R+L	1.89	2.62	2.42	2.32	1.92	1.99

MAN TruckScenes · Vocabulary Top-1 ADE@6s (m) · Lower is better
Sensor input	Clear	Overcast	Rain	Illumination	Twilight	Dark
C	3.77	7.76	3.60	5.87	5.24	5.09
R	1.68	2.52	1.89	2.14	1.99	1.89
L	6.94	18.06	3.89	12.81	2.20	5.78
C+R	1.77	2.75	2.16	2.30	2.01	1.96
C+L	4.11	7.38	3.04	5.60	3.25	7.20
R+L	1.74	2.57	2.14	2.21	1.88	1.97
C+R+L	2.02	3.13	1.27	2.44	2.15	2.52

MAN TruckScenes · Diffusion Top-1 ADE@6s (m) · Lower is better
Sensor input	Clear	Overcast	Rain	Illumination	Twilight	Dark
C	4.40	10.63	2.74	7.01	2.28	8.58
R	1.67	2.67	1.87	2.23	2.05	1.87
L	3.93	8.91	6.58	6.51	5.25	6.21
C+R	1.67	2.80	2.48	2.33	2.01	1.86
C+L	4.04	9.92	3.52	7.06	2.95	7.42
R+L	1.79	2.73	2.66	2.35	2.07	1.91
C+R+L	1.84	2.95	2.40	2.46	1.92	2.10

C = camera · R = 4D radar · L = LiDAR. Bold: lowest error in each condition. Aggregate results from Fig. 4.

Fig. 4 · Original condition-stratified results

Illustrative examples

Illustrative example + aggregate charts DSERT-RoLL · Heavy Snow

4D Radar only · charts: C / R / L The particularly large improvements under adverse weather and challenging illumination indicate that 4D radar is especially valuable when camera or LiDAR observations become less reliable. Left: an illustrative example. Right: condition-specific aggregate results.

JavaScript is off. View the clips directly: DSERT · Heavy Snow ↗ DSERT · Heavy Rain ↗ MAN · Dark ↗ MAN · Rain ↗

Perspective 03 / Multimodal complementarity

Multi-modal Complementarity.

Sensor-configuration comparisons further demonstrate the complementary value of 4D radar to other modalities, while revealing dependencies on sensor combination and planning architecture.

Illustrative example DSERT-RoLL · Heavy Rain

Camera → Camera + 4D Radar · radar-only reference Representative planning predictions and their corresponding BEV trajectories. These are illustrative six-second open-loop predictions, not aggregate evidence.

JavaScript is off. View the clips directly: DSERT · Heavy Rain ↗ DSERT · Light Snow ↗ MAN · Overcast ↗ MAN · Dark ↗

Aggregate Radar Gain

4D Radar as a Planning Modality.

Across the two datasets, radar reduces Top1ADE in 22 of the 24 matched comparisons, corresponding to three base sensor configurations evaluated with four planning baselines on both datasets.

Planner

DSERT-RoLL · Sparse Query

Top-1 ADE@6s (m) ↓ Lower is better

CameraC

2.95

4D RadarR

2.45

LiDARL

2.52

Camera + RadarC+R

2.38

Camera + LiDARC+L

3.24

Radar + LiDARR+L

3.26

Camera + Radar + LiDARC+R+L

3.08

With 4D radarWithout 4D radarAxis fixed across planners within each dataset.

C = camera · R = 4D radar · L = LiDAR

For this configuration

Adding radar to camera reduces ADE from 2.95 to 2.38 m. Adding radar to LiDAR increases it from 2.52 to 3.26 m.

View all 56 results from Table III

DSERT-RoLL · Top-1 ADE@6s (m), lower is better
Sensor input	GRU	Sparse Query	Vocabulary	Diffusion
C	2.71	2.95	3.20	3.06
R	2.55	2.45	2.71	2.51
L	2.71	2.52	2.81	2.53
C+R	2.57	2.38	3.04	2.86
C+L	2.76	3.24	3.06	2.91
R+L	2.49	3.26	2.53	2.41
C+R+L	2.48	3.08	3.37	2.60

MAN TruckScenes · Top-1 ADE@6s (m), lower is better
Sensor input	GRU	Sparse Query	Vocabulary	Diffusion
C	5.54	5.78	5.75	7.03
R	2.10	1.96	2.11	2.18
L	6.23	6.11	11.81	6.43
C+R	2.26	2.36	2.25	2.26
C+L	5.21	4.67	5.70	6.96
R+L	1.76	1.98	2.17	2.29
C+R+L	2.56	2.28	2.44	2.40

C = camera, R = 4D radar, L = LiDAR. Bold indicates the best of all seven configurations for each planner. Values are transcribed from the supplied manuscript; differences of rounded table entries may vary by 0.01 m from reported gains.

8 / 8

Standalone planning capability

Radar-only input achieves the lowest single-modality Top1ADE for every planning baseline on both datasets.

22 / 24

Multi-modal complementarity

Across the two datasets, radar reduces Top1ADE in 22 of the 24 matched comparisons.

19 / 20

Adverse conditions

On DSERT-RoLL, radar-only input outperforms camera-only input in 19 of the 20 adverse-weather baseline-condition pairs.

Modular benchmark design

Plug and Play design for future research

The modular design supports additional datasets, sensing modalities, and planners, providing a flexible foundation for future research on radar-based autonomous driving planning.

01 / Datasets

Dataset adaptation

DSERT-RoLL
MAN TruckScenes

02 / Sensing modalities

Sensor input encoding

Camera · 4D Radar · LiDAR

Seven sensor configurations

03 / Planning methods

Planning-specific scene representation

GRU · Sparse Query
Vocabulary · Diffusion

Standardized interfaces

This design allows individual components to be exchanged independently while maintaining a consistent evaluation protocol across sensor configurations and planning methods.

Resources & citation

Build on Radar2Plan.

The Radar2Plan preprint is publicly available on arXiv. Please use the citation below when referring to the benchmark.

Read the paper on arXiv

arXiv:2610.06121 · First posted 5 October 2026.
Download the paper PDF

@misc{yao2026radar2plan,
  title  = {Radar2Plan: Benchmarking 4D Radar for End-to-End
            Open-Loop Ego-Trajectory Planning},
  author = {Yao, Ling and Xiao, Yichun and Jin, Jin and
            Zhang, Yihan and Ding, Fangqiang},
  year   = {2026},
  eprint = {2610.06121},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url    = {https://arxiv.org/abs/2610.06121}
}

When visibility changes, what can radar still see?

Robust sensing.But robust planning?

Can a planner generate competitive ego trajectories using 4D radar as its only environmental sensing modality?

Under adverse conditions, does radar-based planning achieve lower planning error than camera or LiDAR-based planning?

Does adding 4D radar improve planning when combined with camera, LiDAR, or camera-LiDAR sensor inputs?

A modular benchmark with standardized interfaces.

Consistent task definitions

Seven sensor configurations

Top-1 ADE at six seconds

DSERT-RoLL

MAN TruckScenes

Standalone Planning Capability.

Planning Robustness under Adverse Conditions.

Illustrative examples

Multi-modal Complementarity.

4D Radar as a Planning Modality.

DSERT-RoLL · Sparse Query

Standalone planning capability

Multi-modal complementarity

Adverse conditions

Plug and Play design for future research

Dataset adaptation

Sensor input encoding

Planning-specific scene representation

Build on Radar2Plan.

Robust sensing.
But robust planning?