Code for our paper (https://arxiv.org/abs/1708.04391) on mapping body-affordances for dimensionality reduction of spaces of actions.
The best-documented example is the Hexapod code, which has comments discussing architectural choices of the neural networks and the detailed procedure for training. The reacher code has many common elements and the detailed explanations are not repeated there.
This code trains a model to control an 8 DoF armature in the presence of obstacles in the environment. The model sees these obstacles via an overhead depth camera. The model learns to map the set of reachable points with an approximately uniform control grid.
This code trains a model to control the locomotion of a hexapod robot. The model can direct the robot to walk or run towards points in the plane.
