Robotic Assembly System with Uncertainties Balancing

Industrial robots are extensively used in assembly and production lines due to their effectiveness in executing repetitive tasks. In structured environments, these robots are preprogrammed and achieve a high success rate in performing different tasks. However, they lack flexibility when it comes to handling changes and unforeseen events during their operations because they lack adequate sensing and learning capabilities, especially in precision assembly scenarios. Consequently, robotic assembly operations in semi-structured or unstructured environments still is a significant challenge. The integration of a force sensor into the assembly system improves the robot with the capability to reduce uncertainties by measuring contact forces and moments, thereby enabling refined manipulative actions. However, this compliance control method requires a fine-tuned parameter setting to work. In recent trends, Reinforcement Learning (RL) has shown its effectiveness in the context of uncertainty and variability. RL also has its own limitations: Firstly, the generalization ability for different tasks and assembly scenarios is hard to achieve. Secondly, because of low sample efficiency in the real world, it is time-consuming to acquire enough interactions with the real environment for the training. Thirdly, the gap between simulation and reality prevents the transfer of the policies learned in simulated environments to real environments. This thesis presents the development of a learning-based control method for the industrial robot in the presence of different uncertainties such as contact dynamics modeling, manufacturing inconsistency, and sensor feedback. Firstly, a framework for robotic assembly is developed. Due to the uncertainties associated with this assembly frame, certain tasks are not performed effectively. These uncertainties and modeled by using the probabilistic method and added to the simulation environment. RL-based admittance controller is designed, trained, and tested in simulation and real-world scenarios. Reward designs are proposed for peg-in-hole (PiH) and a belt looping (BL) task. To further increase the performance, Multi-Objective Reinforcement Learning (MORL) is adapted to the admittance controller. The preferences among the multiple objectives could be assigned according to specific application scenarios and the trade-offs between performance objectives associated with the level of uncertainty. The proposed method has been applied to a PiH and a BL task, and several agents with different preferences are trained. In simulated environments, agents trained with the proposed method outperformed those trained with traditional reward design. Moreover, these trained agents are transferred to a physical robotic manipulator, and experiments showed a similar performance compared with the simulation. The contribution of this thesis research is its insights towards the goal of enabling industrial robots a human-like performance in the presence of uncertainties. The proposed method has the potential to increase the efficiency of finding an optimal or near-optimal learned policy in semi-structured environments.