Inverse Kinematics Feed Forward Neural Network

Multi-layer feed-forward neural network for estimating a 6DOF robot's internal joint angles from cartesian coordinates. The neural network is modeled around the concept of how transformation and coordinate matrices feedforward to estimate the angle of each internal joint.

Installation

Requirements:

• python version 3.7.6 or later
• conda version 3.18.11 or later
• torch version 1.7.1 or later, torchvision, and 'pytorch-pfn-extras' To install: git clone github.com/alexander-jh/KineNet.git conda create --name <env_name> --file requirements.txt

Training

Execute command from conda environment. python3 src/train.py

Generating Training Data

python3 src/SimDataGenerator.py <# of batches to generate>.

Testing

Specify type of data to use: python 3 src/test.py

Results

Over the training cycle there's an approximate loss of around 10% on average across all 6 joints.

Looking at results by joint, the T-joint is the only joint that's consistently inaccurate.

Mean difference in radians between joint estimation and actual.

S-Joint	L-Joint	U-Joint	R-Joint	B-Joint
-0.0035	0.0075	0.0042	-0.0129	0.0036