Learning From Sparse Demonstrations

In this article, we develop the method of continuous Pontryagin differentiable programming (Continuous PDP), which enables a robot to learn an objective function from few sparsely demonstrated keyframes. The keyframes, labeled with some time stamps, are desired task-space outputs, is expected follow sequentially. stamps keyframes can be different robot's actual execution. jointly finds and time-warping such that resulting trajectory sequentially follows minimal discrepancy loss. Continuous PDP minimizes loss using projected gradient descent by efficiently solving respect unknown parameters. first evaluated on simulated arm then applied 6-DoF quadrotor for motion planning in unmodeled environments. results show efficiency method, its ability handle misalignment between execution, generalization learning into unseen conditions.