/agribot

Primary LanguagePython

AgriBot: Intelligent Agricultural Robot

demo-image demo-image2

To Do

  • Check if scan or realsense/scan for a real robot and simulated?

History

  • 01/02/2024: Remove fake_localization and the gen_pose option.
  • 12/02/2023: Fix data normalization algorithm.
  • 12/01/2023: Add checkpoint saving option. The default is False.
  • 11/22/2023: Add gen_pose option for fake_localization.
  • 11/22/2023: Add p3d and fake_localization for pose data.
  • 11/20/2023: Merge the noetic-devel branch to master.
  • 11/20/2023: Add scout_control_translator.
  • 11/19/2023: Add joystick_translator.
  • 11/16/2023: Start Noetic version.
  • 11/14/2023: Add several How-To's.
  • 11/14/2023: Add view_drive.py avoiding a direct use of drive_view.py.
  • 11/12/2023: Add test_drive.py replacing a direct use of drive_log.py.
  • 11/11/2023: Test in the latest TensorFlow.
  • 11/11/2023: Change model format to Keras.
  • 11/11/2023: Clean up.
  • 10/31/2023: Start with submodules.

Example Results

The following is a yaml file example for neural network training.

###############################################################################
# This is a jaerock's configuration for scout

# file name: <robot-name>-<person-name-who-trained-network>
# example: if jaerock trained network with data using the robot, scout
#          the file name must be scout-jaerock.yaml

version: 1.00

# history
# 1.00 - start

######################################################
## neural network 

# network definition
network_type: 3        # refer to neural_net/const.py
lstm: False

lstm_timestep: 20
lstm_lr: 0.00001        # 
cnn_lr:  0.001          # default of Adam is 0.001 

# normalize data
normalize_data: True
num_bins: 50
samples_per_bin: 1000

# training 
data_shuffle: True
validation_rate: 0.3
num_epochs: 100
batch_size: 32
num_inputs: 1     # input: image, velocity
num_outputs: 1    # output: steering_angle, throttle
early_stopping_patience: 3
checkpoint: False
# tensorboard log dir
tensorboard_log_dir: logs/scalars/

# data augmentation
data_aug_flip: True
data_aug_bright: True
data_aug_shift: False

# input image size to the neural network
input_image_width:  160
input_image_height: 160
input_image_depth:  3

# steering data preprocessing
# - steering angle adjustment
steering_angle_scale: 1.0
# - steering wheel jitter tolerance
steering_angle_jitter_tolerance: 0.01

With the scout-jaerock-orchard_samples1000.yaml configuration, the training result is as follows.

normalization The total number of bins is 50 and the maximum samples per bin is 1000 for training data normalization.

training The MSE loss curves are nicely smooth. The EarlyStopping callback was used.

scatter The prediction performance around zeo steering angle is overall good. But, it is not particularly good around zero steering. This might be able to be improved by adding more training data, where the robot is moving forward in the center of the orchard path.

How to Clone

This repo has submodules. Please clone with the following command.

git clone --recurse-submodules https://github.com/jrkwon/agribot

TensorFlow Installation with CUDA, cuDNN in Anaconda Env

Additional SDKs

To run agribot, you must install SDKs for sensors and additional packages, including YDLIDAR SDK and Cartographer. The details can be found https://github.com/jrkwon/agribot_ros.

Prior to Use

Create the agribot conda environment.

Note: We tested rospy.init_node() of ROS Noetic with Python 3.11. Well... init_node() never returns. We tried Pytnon 3.10 that works well with rospy.init_node(). Thus, this conda environment has Python 3.10.

conda env create --file config/conda/environment.yaml

How to Use

./start_simul.sh launch_filename(without .launch)

The name should be without the .launch file extension.

If scout_orchard_world is the launch file name (without the .launch extension)that you would like to use, do this.

./start_simul.sh scout_orchard_world

Then, open another terminal. Activate the agribot environment.

conda activate agribot

Then source the setup.bash. You must be at the agribot directory when you do this.

source setup.bash

How to set configuration files

In the config directory, there are config-<name>.yaml files. You can choose which yaml should be used by changing the name CONFIG_FILENAME in neural_net/const.py

How to Collect Data

The default data location is agribot/e2e-dataset/<data-id>. If jaerock collected a data at 5:50:10 PM, Nov 12, 2023, then the directory name of the dataset is agribot/e2e-dataset/jaerock/2023-11-12-17-50-10.

cd path/to/agribot
source setup.bash

rosrun data_collection data_collection.py <data-id>

If you have a separate storage to save data, you can set the path_to_e2e_data, which is a rosparm.

cd path/to/agribot
source setup.bash

rosparam set path_to_e2e_data path/to/location/to/save  
rosrun data_collection data_collection.py <data-id>

The data directory has a lot of images in most cases. So, it can be slow when you open the directory. Users need to open a CSV file inside the directory in many cases. Here is a script that can make a symblic link to the agribot/e2e-dataset/<data-id>.

cd path/to/agribot

sh make_simlink.sh path/to/data-dir

How to Train

Note:

  • All checkpoint files are now saved at the yy-mm-dd-hh-mm-ss_ckpt directory in agribot/e2e-dataset/<data-id>
  • The trained model is saved as the latest Keras model.
  • All pictures and csv files (comprison, err_hist, log, scatter) are generated with the format: <dataset-dir-name>_<config_name>_<trained-model-name>_<timestamp> to identify when the training/driving was done with which dataset, which configuration, and which model.
  • When neural_net related scripts are run, you are expected at the agribot directory.

Here is an example:

conda activate agribot
cd path/to/agribot

source setup.bash
python neural_net/train.py e2e-dataset/<id>/yyy-mm-dd-hh-mm-ss

How to Do Test Drive

Once you have a trained model, you can do a test drive with a dataset.

Let us assume followings as an example.

  • data id: jaerock
  • dataset name: 2023-10-31-14-56-16
  • config name: scout-jaerock-orchard--aug
  • network id: 3
  • timestamp: 2023-11-12-16-25-31-079

Then, a trained model name is 2023-10-31-14-56-16_scout-jaerock-orchard--aug_n3_2023-11-12-16-25-31-079.keras

Let us test the model with a dataset. You may test the trained network with your training dataset for an initial trial. Then, you may test it with a test data set.

conda activate agribot
cd path/to/agribot

source setup.bash
python neural_net/test_drive.py path/to/model path/to/data-dir

How to Test Single Prediction

conda activate agribot
cd path/to/agribot

source setup.bash
python neural_net/test_run.py path/to/model path/to/image

How to Visualize Heatmap

conda activate agribot
cd path/to/agribot

source setup.bash
python neural_net/visualize_heatmap.py path/to/model path/to/image

How to Generate Images and Movie Showing Control Signals

  • Generate images showing control signals.

    conda activate agribot
cd path/to/agribot

source setup.bash
python neural_net/view_drive.py path/to/model path/to/data-dir path/to/target-dir

    The images will be generated under path/to/target-dir/data-dir-name

  • Generate mp4 using the generated images.

    cd path/to/agribot

sh make_video.sh path/to/target-dir/data-dir-name

    Caution: Do not include '/' at the last of the directory name.

How to Let the Trained Neural Network Drive the Robot

The best practice is to

  • Start a Gazebo simulated world without any conda environment activated.

    cd path/to/agribot
source setup.bash

./start_simul.sh <launch name without '.launch'>

  • Open a new terminal.

  • Activate the agribot conda environment

    conda activate agribot

  • Start run_neural.

    rosrun run_neural run_neural.py path/to/trained/model

Troubleshooting

  • On Ubuntu 20.04, cv_bridge may generate this error below.

    Traceback (most recent call last):
File "/opt/ros/noetic/lib/python3/dist-packages/rospy/topics.py", line 750, in _invoke_callback
    cb(msg)
File "/home/jaerock/projects/agribot/catkin_ws/src/run_neural/scripts/run_neural.py", line 67, in _controller_cb
    img = self.ic.imgmsg_to_opencv(image)
File "/home/jaerock/projects/agribot/neural_net/image_converter.py", line 29, in imgmsg_to_opencv
    cv_img = self.bridge.imgmsg_to_cv2(img_msg, 'rgb8')
File "/opt/ros/noetic/lib/python3/dist-packages/cv_bridge/core.py", line 163, in imgmsg_to_cv2
    dtype, n_channels = self.encoding_to_dtype_with_channels(img_msg.encoding)
File "/opt/ros/noetic/lib/python3/dist-packages/cv_bridge/core.py", line 99, in encoding_to_dtype_with_channels
    return self.cvtype2_to_dtype_with_channels(self.encoding_to_cvtype2(encoding))
File "/opt/ros/noetic/lib/python3/dist-packages/cv_bridge/core.py", line 91, in encoding_to_cvtype2
    from cv_bridge.boost.cv_bridge_boost import getCvType
ImportError: /lib/x86_64-linux-gnu/libp11-kit.so.0: undefined symbol: ffi_type_pointer, version LIBFFI_BASE_7.0

  • This is because cv_bridge is dynamically linked with libffi.8 not libffi.7. Remove that library.

    rm ${CONDA_PREFIX}/lib/libffi.7.so ${CONDA_PREFIX}/lib/libffi.so.7

Acknowledgments

System Design and Implementation

  • Jaerock Kwon, PhD, Assistant Professor, Electrical and Computer Engineering, University of Michigan-Dearborn

Implementation and Contributors

  • Elahe Delavari, PhD student, Electrical and Computer Engineering, University of Michigan-Dearborn
  • Feeza Khanzada, PhD student, Electrical and Computer Engineering, University of Michigan-Dearborn