ANN-based-approximate-kinematic-transformations

The "pmptrain.py" file uses ANN to fit the relationship between joint angles and end positions, evaluating the Jacobian matrix via the TF GradientTape API.

The "pmptrain_chain_rule.py" file is an example of using the chain rule to calculate the Jacobian matrix (Not recommended for networks with a high number of layers and nerves).

The "IndhRobot_example.txt" and "OutdhRobot_example.txt" show the format of the training data：

"IndhRobot_example.txt" -- Joint angles (degree)

"OutdhRobot_example.txt" -- End position (mm)

The rows in both files should correspond to each other, where the joint angles correspond to their end positions.

To train an ANN model using the data:

python3 pmptrain.py

Once the training is done, the model will be saved. To transfer or fine-tune the model via transfer learning (make sure the data and model loaded correctly):

python3 transfertrain.py

Passive motion paradigm implementation via deep neural networks

We also provide a Python version of the Passive Motion Paradigm (PMP) motion model, which uses the deep neural networks-based PMP to realize the goal-directed motion:

python3 pmp_ANN.py

Besides, the non-ANN PMP is also available:

python3 pmp_non_ANN.py