Apprenticeship Learning for Robotic Control.

摘要

Our research had three main thrusts: (i) Optimization-based motion planning: Traditional approaches to motion planning tend to slow down in high-dimensional spaces and for curvature constrained problems. We developed an optimization-based approach that can efficiently solve such problems. (ii) Belief space planning: To account for uncertainty about the environment and robot belief space planning tries to find a plan that optimizes (on expectation) the sequence of probability distributions that would result from that plan. As a consequence belief space planning anticipates (and accounts for) sensory information, the expected magnitude of perturbations, and controllability. We made contributions in handling collisions, handling occlusions, scaling up belief space planning, going beyond unimodal beliefs, and hierarchical belief space planning. (iii) Learning from demonstrations: We developed an approach to generalize demonstrated motion in training scenarios to test scenarios. At the core of our approach is non-rigid registration to warp the training scene onto the test scene. While registration is only concerned with the objects and their environment, we show that it is possible to meaningfully extrapolate the registration and to also warp the robot gripper trajectories from training scene to test scene. Our approach is particularly appealing for the manipulation of deformable objects, which present the robot with a very high-dimensional state space and large amounts of variability, making them particularly challenging for robots to manipulate. It has enabled autonomous knot tying for a wide range of knot-types and automation of some simplified suturing tasks.