Job title: Master thesis Context-Aware Active Grasping on Unseen Objects
Job description: Company Description
At Bosch, we shape the future by inventing high-quality technologies and services that spark enthusiasm and enrich people’s lives. Our promise to our associates is rock-solid: we grow together, we enjoy our work, and we inspire each other. Join in and feel the difference.
The Robert Bosch GmbH is looking forward to your application!
Grasping is one of the fundamental and crucial subtasks of robotic manipulation. However, reliable robotic grasping is challenging due to imprecise sensing and actuation. Recently, many works[1, 2] have shown advances in predicting grasp success directly from depth images by training a Deep Convolutional Neural Network (CNN) with a large-scale synthetic dataset. However, these methods focus more on local geometric features of the object due to a single view or partial observation. Also, objects with similar shapes could include different properties, e.g., material, texture, mass distribution and friction coefficient. Therefore, without contexts and using depth, the model cannot capture these global features. Thus, this might lead to a failure in many real-world applications.
In this thesis, we propose to leverage global features from contexts in an active-learning manner using CNPs. The features of an object (e.g., inhomogeneous density, friction coefficient, material or texture of different parts) should be captured implicitly from the collected prior trials. The model is expected to form the likelihood of underlying physical parameters and choose the best grasping point in the next try. Furthermore, we will transfer and test the performance in the real world with collecting trials in an online fashion, though the model is trained fully with simulation. This work could potentially complement our current project, for instance, picking objects with plastic packages or inhomogeneous properties.
- You are expected to finish the following tasks in this thesis:
- You will improve the existing simulation framework with Mujoco, align with the real robot setup.
- Furthermore, you will generate synthetic dataset for training and evaluation.
- You will extend baselines [1, 2] with contextual encoder using CNPs.
- Last but not least, you will benchmark grasping performance in the real world with more challenging objects with Franka robot.
- Education: Master studies in the field of Robotics, Information Technology or comparable
- Personality and Working Practice: innovative, responsible, self-motivated and team-minded
- Experience and Knowledge: good knowledges in Robotics, Grasping, Deep Learning, Meta-Learning; knowledge of Mujoco and contextual coding with CNPs
- Languages: very good in English
Start: according to prior agreement
Duration: 6 months
 Mahler, Jeffrey, Jacky Liang, Sherdil Niyaz, Michael Laskey, Richard Doan, Xinyu Liu, Juan Aparicio Ojea, and Ken Goldberg, “Dex-Net 2.0: Deep Learning to Plan Robust Grasps with Synthetic Point Clouds and Analytic Grasp Metrics,” arXiv preprint arXiv:1703.09312, 2017.
 Sundermeyer, Martin and Mousavian, Arsalan and Triebel, Rudolph and Fox, Dieter, “Contact-GraspNet: Efficient 6-DoF Grasp Generation in Cluttered Scenes,” IEEE International Conference on Robotics and Automation (ICRA), 2021.
 Garnelo M, Rosenbaum D, Maddison CJ, Ramalho T, Saxton D, Shanahan M, Teh YW, Rezende DJ, Eslami SM, “Conditional Neural Processes,” In International Conference on Machine Learning (ICML), 2018.
Requirement for this thesis is the enrollment at university. Please attach a motivation letter, your CV, transcript of records, examination regulations and if indicated a valid work and residence permit.
Need further information about the job?
Ning Gao (Business Department)
+49 172 5254685
Work and Live in Germany? find jobs and information about visa language and applications here Jobs-in-Germany.net