This is a reinforcement learning (-ish) system whereby two dimensional “organisms” comprising rigid body segments and damped springs learn to walk. The project uses Bayesian methods, maximum likelihood, mixture models, backpropagating neural networks, and gradient ascent with a number of popular Python libraries: NumPy, matplotlib, Jupyter, TensorFlow, pymunk, etc. Watch the linked video here for an idea of agent performance after learning by observing data from human-controlled interaction.
I hope to revisit this domain someday.