Script

Usage

Use python SCRIPT_NAME -d FILE_PATH,FILE_PATH,... [OPTIONS] to run the program.

Example Usage

python SVDENSE_Baxter7DOF.py -d ./Baxter_14_22x22x21.npz -l 0.00001

python SVDENSE_Baxter7DOF.py -d ./Baxter_14_22x22x21.npz -s 90000 -p -r ./checkpoints/08102020180642/

Description

The script files provided allow the user to train the networks and predict output using trained network checkpoints.

The scripts provided are:

• SVDENSE_Baxter7DOF.py
• SVDENSE_ClosedLoop9DOF.py
• SVDENSE_Kuka7DOF_101010.py
• SVDENSE_Kuka7DOF_151515.py
• SVDENSE_Kuka7DOF_202020.py
• SVDENSE_Kuka7DOF_252525.py
• SVDENSE_Kuka7DOF_303030.py
• SVDENSE_Mico6DOF.py
• SVDENSE_Planar16DOF.py
• SVDENSE_Prismatic8DOF.py
• SVDENSE_Sawyer8DOF.py
• SVDENSE_UR3_6DOF.py
• SVDENSE_UR5_6DOF.py
• SVDENSE_UR10_6DOF.py

Options

Below are a list of command line parameters:

• -d FILE_PATH,FILE_PATH,... A comma separate list of file paths for the data. The program accepts data in the format decribed below in the Data section. This option is required.
• -C data_limit Limits the number of samples used by the network from the given data files. Default: 0 (All samples)+
• -s skip_count Skip the first skip_count samples in the data file.
• -p Use the network to predict outputs for the given samples and save predictions to a file. Note: This option was used for evaluation purposes and requires label data to generate labels for comparison.
• -c checkpoint_root Path to a directory. The network records the checkpoint data and logs to a time stamped sub directory within checkpoint_root directory. Default: checkpoints
• -r checkpoint_dir The path to restore the network weights.
• -e epochs The maximum number of epochs for training. Default 1000
• -b batch_size The network training batch size. Default: 100.
• -i tolerance The maximum number of epochs for which the network will train without new improvements. Default 100
• -l learning_rate The learning rate parameter for the training optimizer. Default 0.1

Data

Description

The files provided are saved scipy.sparse.csr_matrix objects. The first Num-Features columns are the start configuration followed by end configuration joint values for the robot. The remaining columns are label values for the Label-Dims volume grid space converted from xyz coorinates to a linear order with x being the most significant axis and z the least significant.

File name format: Robot_Num-Features_Label-Dims.npz

• Baxter_14_22x22x21.npz
• ClosedLoop_18_41x50x3.npz
• Kuka_14_10x10x10.npz
• Kuka_14_15x15x15.npz
• Kuka_14_20x20x20.npz
• Kuka_14_25x25x25.npz
• Kuka_14_30x30x30.npz
• Mico_12_15x15x15.npz
• Planar_32_1x100x100.npz
• Prismatic_16_88x88x2.npz
• Sawyer_16_26x26x26.npz
• UR3_12_15x15x15.npz
• UR5_12_20x20x20.npz
• UR10_12_30x30x30.npz

Software and Libraries

The following are a list of software and libraries used for the project:

• Python 3.6.9
• numpy 1.16.3
• scipy 0.19.1
• tensorflow-gpu 1.14
• tensorflow-estimator 1.14