C-LEARN: Learning Geometric Constraints from Demonstrations for Multi-Step Manipulation in Shared Autonomy
Not totally sure on the technical details for this, but at a high level it builds a knowledge base for robot manipulation with constraints, which can be enforced (with certainty) unlike with an LfD using supervised learning. So, augment LfD with support for geometric constraints.
Like a lot of meta-learning now (from Berkeley and OpenAI, at least, heh) they do a two stage process where they build a knowledge base (the prior) and then learns a multi-step manipulation task using a single demonstration.
The other cool thing is the robot-to-robot transfer despite differing kinematics.