Image-Based Robot Learning for Manipulation

In this work, I surveyed the different ways robot learning can be conducted. I focused on comparing deep learning approaches with the normal pipelined machine learning solutions. I, Further, experimented and evaluated the performance of two different deep learning solutions with imitation learning, using Panda robot arm for the task of reaching a target object in the scene. My experiments show that key-points based CNNs solution (with the special spatial softmax layer) outperforms the direct CNNs one (normal CNN with flattened layer) with a success rate of 80.40%.

Illusterative video:

The target object is the red sphere, while the other two are random distractors