Many robot manipulation tasks require planning robot motions relative to specific landmarks in a scene. Using the approach of Demonstration-Guided Motion Planning (DGMP), the robot uses human-guided demonstrations to automatically learn task-relevant landmarks and the motion needed to accomplish the task relative to those landmarks. Using our new approach, the Baxter robot learns to autonomously pour liquids and transfer powder using a spoon with less human effort required than in prior methods, even in an environment with new obstacles and when people move task-relevant objects in the scene.
Citation:
Chris Bowen and Ron Alterovitz, "Asymptotically Optimal Motion Planning for Tasks Using Learned Virtual Landmarks," IEEE Robotics and Automation Letters, 2016.
