The network’s architecture

ff net
Fig. 1: The network’s architecture for learning inverse kinematics.

A simple fully connected network with one hidden layer and 128 units is used to produce joint coordinates $\mathbf{y}$ in joint space, given the tool center point’s position $\mathbf{x}$ in cartesian space.

10 epochs

Fig. 2: The robot’s attempt to follow a circle after 10 epochs of training.

100 epochs

Fig. 3: The robot’s attempt to follow a circle after 100 epochs of training.

1,000 epochs

Fig. 4: The robot’s attempt to follow a circle after 1000 epochs of training.

100,000 epochs

Fig. 5: The robot’s attempt to follow a circle after 100,000 epochs of training.