DPNet: Doppler LiDAR Motion Planning for Highly-Dynamic Environments

Project Page of The RA-L '26 Paper DPNet

Wei Zuo, Zeyi Ren, Chengyang Li, Yikun Wang, Mingle Zhao, Shuai Wang, Wei Sui, Fei Gao, Yik-Chung Wu, and Chengzhong Xu

Table of Contents

• 🚀 Major Updates
• 📖 Introduction
• 🛠️ Prerequisite
• 🔌 Usage
  • Quick Install & First Run
  • Train D-KalmanNet
• 💭 Q & A
• 📋 Miscellaneous
• 🙌 Citation
• 📪 Acknowledgement
• 📇 License

🚀 Major Updates

[04-Jun-2026] Paper page on Hugging Face released here.
[21-May-2026] Code for D-KalmanNet training is released.
[28-Apr-2026] Code is released.
[10-Apr-2026] Acceptance to IEEE Robotics and Automation Letters (RA-L).

📖 Introduction

Based on cutting-edge technologies like Frequency-Modulated Continuous-Wave (FMCW), Doppler LiDAR is a powerful sensor that provides the 4-th measuring dimension: Doppler velocity. Apart from traditional spatial (x, y, z) measurement, Doppler velocity directly captures the instantaneous spatio-temporal knowledge at per-point level, thus allowing direct, granular, and accurate 4D scene understanding. DPNet is the first framework to integrate Doppler LiDAR into closed-loop motion planning. By incorporating and leveraging Doppler LiDAR, DPNet introduces:

• Doppler Kalman Neural Network (D-KalmanNet): a real-time obstacle motion prediction module.

• Doppler-Tuned Model Predictive Control (DT-MPC): a runtime prediction-triggered controller tuning module.

Built upon these two modules, DPNet achieves agile collision-free motion planning among fast moving obstacles, opening up new research paradigms of Doppler LiDAR guided robot motion planning and autonomous driving.

🔥 Video intro available on YouTube or bilibili

🏎️ Experimental demonstration on random testbed

dpnet_on_dynabarn.mp4

🛠️ Prerequisite

Step 1 - ROS Noetic

DPNet requires ROS-noetic. Migration to ROS 2 is in our future plan.

Step 2 - carla-aeva

DPNet requires the carla-aeva, where Aeva, Inc enhances carla by Doppler LiDAR integration. To obtain, there are two ways:

• Build from Source: Follow the steps in carla-aeva to build from source, optionally adding your custom items.
• Quick Download: We provide pre-compiled version, available at Google Drive or Baidu NetDisk.

Step 3 - Python API for carla-aeva

In Step 2, if you choose

• Build from Souce: Now you can compile your own Python API for any dedicated Python version.
• Quick Download: Now you will be using the provided Python API Release for quick start.

Step 4 - AevaScenes

Note

You can skip this step if you aim at a quick test of DPNet's navigation in highly-dynamic envinronments.

• To train D-KalmanNet, you should download the real-world Doppler LiDAR dataset AevaScenes.

🔌 Usage

Quick Install & First Run

• We encourage using Conda to manage Python environments. Ensure that ROS-noetic and carla-aeva are installed.
• Launch Carla first:

cd path_to_carla_aeva_on_your_computer
./CarlaUE4.sh

• Then, run DPNet in a new terminal:

# clone DPNet with ros-bridge-DopplerLiDAR submodule
cd ~
git clone --recurse-submodules https://github.com/UUwei-zuo/DPNet.git
cd DPNet

# create Conda venv 
conda create -n dpnet python=3.9 -y
conda activate dpnet
pip install -r requirements.txt
conda install libffi==3.3 -y

# API for more Python versions are provided in project release
# If you build from source, you can install any dedicated API version
WHEEL_URL_39="https://github.com/UUwei-zuo/DPNet/releases/download/v1.0.0/carla-0.9.15-cp39-cp39-linux_x86_64.whl"
WHEEL_FILE="carla-0.9.15-cp39-cp39-linux_x86_64.whl"
wget $WHEEL_URL_39
pip install $WHEEL_FILE && rm $WHEEL_FILE

# build workspace
source /opt/ros/noetic/setup.bash
catkin_make
source devel/setup.bash

# run DPNet example
python3 ./examples/DPNet_run.py

Configuration

• Highly-Dynamic Testbed

To test DPNet, scripts/carla_dynabarn.py is provided to completely randomize highly-dynamic obstacles following DynaBARN. The Basic setting is at examples/DPNet_hyperparameters.yaml, where you can conveniently configure obstacle number/seed setting with:

Carla_DynaBARN:
  vehicles: 3 # obstacle number
  seed: 79 # obstacle seed

Advanced setting is at scripts/carla_dynabarn.py.

• DPNet Hyperparameters

Basic hyperparameter settings are at examples/DPNet_hyperparameters.yaml. Detailed settings can be found in src/dpnet_planner_ros/src/DPNet_planner.py, src/dpnet_planner_ros/src/DTMPC.py, and src/dpnet_planner_ros/src/action_solver.py.

• Doppler LiDAR Scan Pattern

carla-aeva uses a scan pattern file to configure Doppler LiDAR simulation. The simulation consumes GPU at runtime, so you can configure the pattern file located at examples/Doppler_ScanPatterns.yaml according to your computer specs. For example, for higher processing frequency, you could reduce points_per_line; for higher perception granularity, you could increase elevation_steps_deg, etc.

Train D-KalmanNet

• Ensure that AevaScenes is downloaded.
• Configure dataset path in examples/d_kalmannet/config.yaml:

Dataset:
  path: # path_to_data_on_your_machine
        # this path usually contains metadata.json

• Then, to train D-KalmanNet:

cd ~/DPNet/examples/d_kalmannet
conda activate dpnet
python train.py

Configuration

• examples/d_kalmannet/config.yaml provides easy access to different training paramters. For example, you can configure:

Dataset:
  split_ratio: [0.8, 0.1, 0.1] # train/val/test split after selection
  seed: 79 # seed for sequence selection, split, and training
Model:
  dt: 0.2 # prediction time step, supporting: 1.0, 0.5, 0.2, 0.1 
Preprocess:
  num_workers: 16 # adjust preprocessing workers according to your CPU specs  

Usage for DPNet

• Weights can be used for DPNet by replacing src/dpnet_planner_ros/src/d_kalmanet/weights/model.pt with your training outputs.

Important

Ensure that dt in examples/DPNet_hyperparameters.yaml is consistent with that of the weight you use.

Run Test-Split Evaluation

• Test-split evaluation is triggered after running a training by default. Optionally, a standalone test evaluator can be launched by:

cd ~/DPNet/examples/d_kalmannet
conda activate dpnet
python run_test.py

Important

Ensure that dt in examples/DPNet_hyperparameters.yaml is consistent with that of the weight you use.

💭 Q & A

• Q1. Why can D-KalmanNet's step interval dt mismatch DPNet planner's horizon interval sample_time in examples/DPNet_hyperparameters.yaml ?
  A1. D-KalmanNet predicts obstacle motions with a velocity-based motion model, e.g., x'=x+vt+0.5*a*t^{2} under constant acceleration. In contrast, carla simulation realizes vehicle motions with a throttle-based motion model. Consequently, adjusting dt and sample_time independently can mitigate the velocity-to-throttle inconsistency.

📋 Miscellaneous

ROS-Bridge for Doppler LiDAR

To connect carla-aeva with ROS, ros-bridge-DopplerLiDAR is provided. It has been included in DPNet as a default submodule.

Future Release Plan

• Training code for D-KalmanNet
• Improved code cleanups

🙌 Citation

We sincerely appreciate your star and citation if you find this work insightful:

@article{zuo2026dpnet,
  author={Zuo, Wei and Ren, Zeyi and Li, Chengyang and Wang, Yikun and Zhao, Mingle and Wang, Shuai and Sui, Wei and Gao, Fei and Wu, Yik-Chung and Xu, Chengzhong},
  journal={IEEE Robotics and Automation Letters}, 
  title={DPNet: Doppler LiDAR Motion Planning for Highly-Dynamic Environments}, 
  year={2026},
  volume={11},
  number={6},
  pages={7190-7197},
  doi={10.1109/LRA.2026.3685933}
}

📪 Acknowledgement

Contact Wei Zuo by zuowei@eee.hku.hk if you have any questions or suggestions.

📇 License

This repository is released under the MIT license. See LICENSE for additional details.