Within the framework of the LEMURS project, Task T2.2 – ML-based Algorithm Evaluation for Manipulation has been successfully carried out through two Master’s theses conducted using an industrial manipulator (UR3e).
This task aims to evaluate the performance of Deep Reinforcement Learning (DRL) and Imitation Learning (IL) algorithms for robotic manipulation, as well as to investigate strategies for transferring learned policies from simulation to real robotic platforms (Sim2Real). In line with the objectives of T2.2, relevant state-of-the-art algorithms were identified, implemented, and experimentally validated on representative industrial manipulation tasks. These include peg-in-hole insertion using DRL approaches, and vision-only assembly using multi-camera feedback through an IL framework based on behavior cloning.
The work has been developed in the context of the Erasmus Mundus Joint Master Degree in Intelligent Field Robotic Systems (IFRoS) and resulted in the following theses.
Supervised Learning for Robot Manipulation
Author: Tanakrit Lertcompeesin
Supervisors: Narcís Palomeras and Hayat RajaniThis thesis explores imitation learning approaches for robotic manipulation, focusing on vision-based policies and industrial assembly tasks.
Deep Reinforcement Learning for Robot Manipulation
Author: Vania Katherine Mulia
Supervisor: Narcís PalomerasThis work evaluates Deep Reinforcement Learning methods for manipulation tasks using the UR3e platform, with experimental validation on peg-in-hole and assembly benchmarks.
Both theses are available in the IFRoS 2025 Proceedings (2023–2025 cohort).