ILQR implementation (following Tassa et al - IROS 1012) for gym pendulum environment, using both known model and linear Gaussian dynamic model learning (Levine et al - JMLR 2016).
-
gaussian_model_learner.ipynb shows how to learn a conditionally linear Gaussian dynamics model by fitting a joint Gaussian to state action pairs.
-
ilqr.ipynb is an ILQR implementation for a kinematic motion model of a wheeled mobile robot.
-
ilqr-gym.ipynb is an ILQR implementation for an OpenAI gym inverted pendulum, given a known dynamics model.
-
ilqr_model_learner.ipynb is an ILQR implementation for an OpenAI gym inverted pendulum, which attempts to learn the dynamics using motor babbling to fit a set of linear Gaussian dynamics models, as in the guided policy search paper. This works pretty poorly, as local linear models make it hard to plan ahead for a longer horizon, and it can be hard to gather data in terminal pendulum positions.