User imitation for multi-purpose robotics

点击翻译按钮获取中文摘要

The objective of the proposed research program is to develop techniques for the programming of multi-purpose agents such as robots using imitation techniques. Imitation, a form of programming by demonstration, is particularly suitable when the user does not have any programming expertise: during "record" mode, the user would simply remote control the robot, while the behavior is being recorded. The recorded logs are stored in a high-level, symbolic form. During "play" mode, the robot uses the recorded traces to autonomously reproduce the behavior that was observed. What is particularly innovative in our approach is that we want to eliminate the need for the user to explicitly encode the goal of the task, as is required in typical case-based reasoning tasks or approaches relying on reinforcement learning. While this limits the ability of the robot to learn from its mistakes and adjust, it also minimizes the amount of input and programming skills required from the user and it allows for the robot to be truly multi-purpose and easily reprogrammable for a different task. Further, building on our initial success with reproducing reactive behavior in simulated environments (such as RoboCup), we propose to expand and generalize our approach to the imitation of "stateful", context-aware behavior in real life robots. This should make the approach particularly suitable for robots to be used in assistive tasks for the elderly or dangerous rescue missions, where the environment is complex, the specific nature of the tasks can often only be determined "on the field" and non-experts will be the ones re-purposing the robot for new tasks.