Robots are becoming increasingly common in modern industries, performing diverse tasks in areas such as manufacturing, e-commerce and even medicine and healthcare. However, robots are powerful only when the task is repetitive and well defined – generalized robotic manipulation is still a very complex operation, since in many realistic cases object shapes and poses are unknown in advance. This is challenging even to powerful recent methods such as reinforcement learning, as interacting with the environment in a real-world robot setting is slow and expensive. Therefore, it is crucial to construct the task and training framework carefully.
In this project, we will design and build a pipeline for robotic pushing tasks, paving the way for a robotic system that can learn general object manipulation. This includes defining the task by deciding what the robot can observe and which actions it can take; planning a heuristic to enable efficient data collection; and training a reinforcement learning algorithm to use the collected data and generalize to new objects.