Robotics has the capability to enhance human abilities in unparalleled
ways. The integration of advanced sensors and precise actuators has
established the foundation for robots to attain these capabilities, as
evidenced by their presence in factories and, in limited cases, in home
settings. Nevertheless, in unpredictable environments, robots struggle to
adjust to new scenarios, collaborate with humans, and are unable to plan
for complex tasks. In this presentation, I will present recent progress
toward bringing autonomous robotic manipulation agents into the real
world by integrating classical planning techniques with modern learning
methods. After a brief introduction to the fundamental concepts of
classical motion planning for robotic manipulators, this talk will present
(a) recent work on how learning methods can enhance the efficiency of
classical motion planners (b) an investigation of robust planning with
inaccurate models through learning, and (c) an exploration of learning
representations for direct image-based planning algorithms.