Chainer

Akihiko Yamaguchi

Chainer (http://chainer.org/) is an easy-to-use implementation of neural networks, implemented in Python. If you want to start a deep learning thing, Chainer is a good starting point.

Here is an example of learning forward dynamics of 7-DoF (degrees of freedom) robot arm:

Deep learning for forward kinematics of 7-DoF arm (testing)
Learning forward kinematics (joint angles q1,...,q7 to end-effector pose x,y,z,qx,qy,qz,qw), with deep neural networks. Green box denotes an analytically computed pose (i.e. target), and purple box denotes an estimated pose with NNs.
Setup:
NNs:  6 hidden layers, 200 units for each.
Activation function: ReLU.
Loss (error/cost/objective) function: mean square error.
# of samples: 20k and 100k.
# of mini batches is 20.
# of training epochs: 100 and 200.
Hack: normalization of quaternion (qx,qy,qz,qw) is needed
when doing estimation.
Implementation: Chainer ( http://chainer.org/ ).