Evaluation of a Pipeline to generate Deep Learning Agents for Fine Motoric Motion Tasks in the Domain of Serial Manipulators

This thesis explores a learning pipeline that trains robotic agents using imitation learning and imitation-bootstrapped reinforcement learning to perform contact-rich manipulation tasks. While imitation learning provides sample-efficient policy initialization from expert demonstrations, it often lacks robustness in unseen situations. Reinforcement learning, on the other hand, provides better generalization due to parallel training but requires extensive interaction time and a well-shaped reward function, making it impractical for realworld training from scratch. To address these limitations, the thesis evaluates a pipeline in which expert demonstrations are first collected in a simulation environment to train the imitation learning agent. This trained imitation learning agent is then used to initialize the policy weights for reinforcement learning training. Additional demonstrations are synthetically augmented to enhance data diversity. The trained model in simulation is validated on national institute of standard and Technology (NIST) and RoboEval benchmarks. Manipulation tasks peg-insertion and plugsocket-insertion are used to validate the results, which demonstrate that the combined approach of imitation learning + reinforcement learning improves the performance of the agent. This work highlights the potential of hybrid learning strategies for robotic skill acquisition.