الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 130 |  |  |
| | | from | 130 | | |
Abstract
Robots are now widely used in industry to perform assembly tasks. However, they do repetitive tasks in structured environments and require a lot of tuning effort. In order to be widely used in industry, robots must be able to perform assembly tasks in new environments using their own sensors.
Assembly tasks in unknown environments is still an open problem, model based control methods require a lot of human effort to be tuned and calibrated and learning based methods are difficult to be applied on real robots due to sample complexity and safety issues during training. In this work, we provide a learning based method that uses meta reinforcement learning to successfully perform peg-in-hole assembly tasks with high values of uncertainty in the hole position and orientation and with grasping errors and poor performance robot controllers. The proposed method can learn to perform the task without human intervention and it can quickly and robustly adapt to new tasks using few sampled data.
The work in this thesis provides multiple ways to solve the main problem in peg-in-hole assembly tasks which is the alignment error between the peg and the hole which causes task failure and produces high contact forces. The proposed multiple methods can be applied using different sensors to handle different sources of uncertainties.
The work in this thesis mainly depends on a meta reinforcement learning algorithm called PEARL (Probabilistic Embeddings for Actor Critic Reinforcement Learning). The current work proposes modifications to the implementation of PEARL algorithm to be more applicable on real robots and to improve the training and adaptation efficiency of the agent. The results in this work show that the proposed modifications improve the training efficiency of the agent by a factor of three in assembly tasks with uncertainty in the hole position, and allows the agent to perform assembly tasks with uncertainty in the hole position and orientation which is unachievable using the original PEARL algorithm. In addition to that, the work proposes an extra supervised learning based procedure that can adapt the meta trained agent to use different types of sensors to perform the assembly tasks using a limited amount of training data and achieves the same performance as the meta trained agent.
According to our knowledge, this is the first work that uses the PEARL meta reinforce- ment learning algorithm to handle uncertainty in the hole orientation around a certain axis and to handle the grasping errors that causes error in the estimated peg position and causes task failure.