L2Calib

Official code of $SE(3)$-Manifold Reinforcement Learning for Robust Extrinsic Calibration with Degenerate Motion Resilience.

L2Calib (Learn-To-Calibrate) is a reinforcement learning-based extrinsic calibration method that currently supports LiDAR-IMU extrinsic calibration, achieving comparable or even better performance than traditional calibration approaches.

Quick Start

dependencies

• Ceres 1.14
• Sophus 1.22

1. Clone the Repository.

cd
git clone https://github.com/APRIL-ZJU/learn-to-calibrate.git
git submodule update --init --recursive # for Traj-LO. Note that if you already have the reference trajectory, you don't need this component

2. Create Conda Environment

conda create -n L2Calib python=3.11
conda activate L2Calib
pip install numpy==1.26.4 torch==2.4.0 empy==3.3.4 pybind11==2.13.1

3. Build RL Environment

cd Environment
export CMAKE_PREFIX_PATH="$(python -m pybind11 --cmakedir)":$CMAKE_PREFIX_PATH
zsh build.sh

4. Before running, make sure the PATHs are specified

export PYTHONPATH=${HOME}/learn-to-calibrate/Environment/build/app:$PYTHONPATH
export PYTHONPATH=${HOME}/learn-to-calibrate/rl_solver:$PYTHONPATH

test with our CSC example

1. Download our handheld rosbag csc_01.bag

2. Replace the bag_dir variable in calib_csc.sh

3. Run zsh demo/calib_csc.sh

test with slam dataset

1. Download NTU VIRAL or MCD VIRAL
2. Replace the bag_dir variable in calib_ntu.sh / calib_mcd.sh
3. Run zsh demo/calib_ntu.sh or zsh demo/calib_mcd.sh

calibrate your own sensors

1. Put your rosbag in an empty folder
2. Specify the fastlio_config,bag_dir,lidar_type,imu_topic and rough_trans in demo/calib.sh.

• rough_trans represents the rough estimating of the translation between the sensors, where 0.1 means 10 cm.
• Note that if you already have the reference IMU trajectory, e.g., the IMU ground truth trajectory obtained from Mocap/RTK systems, you can set use_imu to True. And ignore the steps below, just run

python train.py --lio-config {FASTLIO CONFIG} --bag-dir {BAG DIR} --alg ppo --SO3-distribution Bingham --num-epochs 8000 --min -{ROUGH} --max 0.1

3. Specify the Traj-LO configuration under Environment/Traj-LO/data/. Only the topic needs to be specified.
4. Run zsh demo/calib.sh

Acknowledegment

Thanks for these awesome works Traj-LO(Lidar-only odometry) Faster-lio (Tightly-coupled LIO) BPP (Bingham policy parameterization for RL)

The parallel environment is adapted from Fast-Evo (Faster evo-ape/traj implemented in C++20)

Thanks Chengrui Zhu for implementing the PPO RL algorithm.