Les commentaires sont clos.

Postdoc or Research Engineer position in Vision and Machine Learning: 6D Pose Estimation for Manipulating Heavy and Complex Geometric Objects with a Robotic Arm

6 Novembre 2023

Catégorie : Post-doctorant

Informations ________________________________________________

Laboratory: LIRIS (UMR 5205)

Type of working contract : CDD

Working contract duration : adjustable between 12 and 24 months

Job starting date : As soon as possible

Job status : full-time

Gross monthly salary : between 2512€ à 2978€ depending on the diploma and experience

Required diploma : PhD diploma for the post-doc position, Master 2 for the Research Engineer position

Place of work : Ecully campus

Deadline for application : December 20, 2023


Research field _____________________________________________

ECL and Laboratory presentation

Founded in 1857, École Centrale de Lyon is one of the top 10 engineering schools in France. It trains more than 3,000 students of 50 different nationalities on its campuses in Écully and Saint-Étienne (ENISE, in-house school): general engineers, specialized engineers, masters and doctoral students. With the Groupe des Écoles Centrale, it has three international locations. The training provided benefits from the excellence of the research carried out in the 6 CNRS-accredited laboratories on its campuses, the 2 international laboratories, the 6 international research networks and the 10 joint laboratories with companies. Its excellent research and high-level teaching have enabled it to establish double degree agreements with prestigious universities and advanced partnerships with numerous companies. With its focus on sobriety, energy, the environment and decarbonization, Centrale Lyon intends to respond to the problems faced by socio-economic players in the major transitions.

This research project will be conducted within the LIRIS Laboratory (UMR 5205), specifically within the Imagine team, whose work aims to develop new learning and vision models for the analysis and understanding of visual data such as images and videos.


Research field presentation :

This mission is part of the acROBaTTH project, which involves collaboration between LIRIS, SETFORGE, and INNOVTEC companies. The project aims to propose new technologies for the robotization of the thermal treatment process for forged parts. Indeed, automating certain handling operations currently performed by human operators will help reduce the labor intensity of the work and prevent accidents and musculoskeletal disorders.

Thus, the overall objective of the mission will be to develop new vision and machine learning methods that will allow, based on camera images, the identification of the poses of objects arranged loosely in bins, so that the robotic arm can safely grasp them, and then determine the optimal placement position for depositing these objects on a tray in order to optimize their arrangement. It should be noted that this mission will focus on image analysis for predicting the grasping position, while the trajectory of the robotic arm will be handled by Innovtec.


Description of the activities

The problem of estimating the pose of objects (position and orientation in a three-dimensional space) is at the heart of this mission. There are multiple scientific challenges involved. Firstly, the scene images captured by the cameras will be highly noisy due to the significant heat conditions and the presence of calamine dust on the surface of the forged parts, which gets released into the air during their manipulation. Moreover, the objects to be grasped can be very heavy (up to 250 kg) and have complex shapes, requiring extremely precise pose estimation to avoid any drops. Lastly, based on the knowledge of the object's pose, the most appropriate gripping point for the robotic arm must be determined to enable the deposition of the object in a suitable position, optimizing the placement of the parts on the thermal treatment pallet.

The preferred approach is to consider a model-based pose estimation method (a 3D model of each part is available). The positions for possible grasping will be determined in advance by an expert human operator for each part model. The objective will be to estimate the pose of the parts in the bulk and determine, at each unpacking step, the most graspable piece (the one least covered by other parts). To achieve this, it will be considered to adapt state-of-the-art solutions for object pose estimation, such as PVNet [Peng19], DenseFusion [Wang19], G2L-Net [Chen20], FS-Net [Chen21] or GDR-Net [Wang21], to the context of bulk handling and partial occlusion [Grard20]. These methods will need to be studied and evaluated to propose an approach suitable for the noisy context and capable of achieving high precision. Furthermore, a continuous learning improvement should also be considered so that new part models can be accommodated as they are introduced into the catalog.



[Grard20] Matthieu Grard, Emmanuel Dellandréa, Liming Chen, "Deep Multicameral Decoding for Localizing Unoccluded Object Instances from a Single RGB Image", International Journal of Computer Vision, vol. 128(5), pp. 1331-1359, 2020.

[Peng19] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, Hujun Bao, "PVNet: Pixel-Wise Voting Network for 6DoF Pose Estimation", Conference on Computer Vision and Pattern Recognition (CVPR), 2019.

[Wang19] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martín-Martín, Cewu Lu, Li Fei-Fei, Silvio Savarese, "DenseFusion: 6D Object Pose Estimation by Iterative Dense Fusion", Conference on Computer Vision and Pattern Recognition (CVPR), 2019.

[Chen20] Wei Chen, Xi Jia, Hyung Jin Chang, Jinming Duan, and Ales Leonardis, "G2L-Net: Global to Local Network for Real-Time 6D Pose Estimation with Embedding Vector Features", Conference on Computer Vision and Pattern Recognition (CVPR), 2020.

[Chen21] Wei Chen, Xi Jia, Hyung Jin Chang, Jinming Duan, Linlin Shen, Ales Leonardis, "FS-Net: Fast Shape-based Network for Category-Level 6D Object Pose Estimation with Decoupled Rotation Mechanism", Conference on Computer Vision and Pattern Recognition (CVPR), 2021.

[Wang21] Gu Wang, Fabian Manhardt, Federico Tombari, Xiangyang Ji,"GDR-Net: Geometry-Guided Direct Regression Network for Monocular 6D Object Pose Estimation", Conference on Computer Vision and Pattern Recognition (CVPR), 2021.


Required skills / qualifications________________________________

Diplomas : PhD diploma for the post-doc position, Master 2 for the Research Engineer position

Knowledge required: Computer Vision, Machine Learning, Deep Learning

Operational skills : Good proficiency in the Python language and deep learning libraries such as PyTorch, proficiency in English, and good written and oral communication skills.

Behavioural skills : Motivation, rigor, autonomy, proactivity.


Recruitment process_______________________________

The recruitment process takes place in two stages, supervised by a recruitment committee, in accordance with Centrale Lyon's OTMR policy.

- Study of the written application: CV + cover letter + PhD diploma

- Selection interview: in person or by videoconference


Deadline for application : December 20, 2023


How to apply________________________________________________

Emmanuel Dellandréa, Associate Professor,