Motion Planning with Success Judgement Model Based on Learning from Demonstration

A technique named Learning from Demonstration allows robots to learn actions in a human living environment from the demonstrations directly. In a learning method from demonstrations directly, however, teaching actions cannot be reused between situations with different restrictions. In this study, we propose a method for training a success judgment model based on Learning from Demonstration and use this as a differentiable loss function of tasks. By formulating the constraints of the action in a manner in mathematical optimization and combining these constraints with the learned success judgment model into a loss function, an action generation model can be trained by the gradient method. This system was verified with the action of scooping up a pancake.