Generative adversarial training of product of policies for robust and adaptive movement primitives

• DOI: 10.48550/arxiv.2011.03316
• Publication Date: 2020-01-01
• Authors: Emmanuel Pignat; Hakan Girgin; Sylvain Calinon
• Keywords: Robustness; Adaptability

In learning from demonstrations, many generative models of trajectories make simplifying assumptions independence. Correctness is sacrificed in the name tractability and speed phase. The ignored dependencies, which often are kinematic dynamic constraints system, then only restored when synthesizing motion, introduces possibly heavy distortions. this work, we propose to use those approximate trajectory distributions as close-to-optimal discriminators popular adversarial framework stabilize accelerate procedure. two problems adaptability robustness addressed with our method. order adapt motions varying contexts, a product Gaussian policies defined several parametrized task spaces. Robustness perturbations dynamics ensured stochastic gradient descent ensemble methods learn dynamics. Two experiments performed on 7-DoF manipulator validate approach.