| KEYWORD |
Generalizing Deep Reinforcement Learning for multi-DoF Robotic Grasping Across Objects
keywords GRASPING AND DEXTEROUS MANIPULATION, HUMANOID ROBOTICS, MACHINE LEARNING, REINFORCEMENT LEARNING, ROBOT LEARNING, ROBOTICS
Reference persons RAFFAELLO CAMORIANO
External reference persons Lorenzo Natale – Istituto Italiano di Tecnologia, Genova
Elisa Maiettini – Istituto Italiano di Tecnologia, Genova
Federico Ceola – Istituto Italiano di Tecnologia, Genova
Research Groups DAUIN - GR-23 - VANDAL - Visual and Multimodal Applied Learning Lab, Humanoid Sensing and Perception (HSP) Group, Istituto Italiano di Tecnologia
Thesis type EXPERIMENTAL, RESEARCH
Description The goal of this thesis is to investigate efficient model-free Deep Reinforcement Learning (DRL) methods for learning to grasp multiple object categories with multi-DoF robotic hands. In particular, the candidate will investigate and develop state-of-the-art DRL methods for training multimodal grasping policies capable of generalizing across different object geometries, only given 2D visual input and tactile information. The RL agent will be jointly trained on multiple object categories, exploring the capabilities of various DRL algorithms and multimodal object representations, with a specific focus on improving data efficiency and generalization. The target robotic platform is the iCub humanoid robot with its dexterous multi-DoF hand, resulting in a 20-dimensional action space.
See the attached thesis proposal description PDF for more details.
See also rl_icub_grasping_thesis_polito_iit.pdf https://drive.google.com/file/d/1Oiinsrs_9ljlmw5TPFeEg76YVE4yn_p3/view?usp=share_link
Required skills - Good proactivity, teamwork experience, and communication skills are a must;
- Excellent programming and software engineering skills are required;
- Strong experience with Python is required, preferably including state-of-the-art ML, RL, data manipulation, and visualization libraries (e.g., PyTorch, OpenAI Gym, RLLib, stable baselines, mushroom, …);
- Experience or strong motivation in working with robotics simulators (e.g. Mujoco,
Bullet, Gazebo, …), and potentially with advanced humanoid robots, are welcome;
- Experience with version control (GIT) and experiment management and visualization
software (e.g., WandB, TensorBoard, …) are a plus;
- Foundational knowledge of the Reinforcement learning paradigm is either expected or
needs to be gained prior to the start of the thesis;
- Knowledge of robotics fundamentals and/or hands-on experience are a plus.
Notes Expected load: Full-time, covering a minimum of 6 months, split in:
- Literature and study – 20%
- Implementation – 40%
- Experiments – 40%
Organization:
- Most activities can be carried out in hybrid mode. The candidate will have access to the VANDAL laboratory premises, with weekly update meetings;
- In-person activities and working sessions at the Center for Robotics and Intelligent Systems of Istituto Italiano di Tecnologia in Genoa, where the HSP research group and the iCub robot are physically located, will also be possible. Potential longer stays are optional and conditioned on available resources and interest.
Deadline 05/04/2025
PROPONI LA TUA CANDIDATURA