Internship : Manipulation using 3D foundation models F/M

About NAVER LABS Europe and Sorbonne University

NAVER LABS Europe is part of the R&D division of NAVER, Korea’s leading Internet portal and a global tech company with a range of services that include search, commerce, content, fintech, robotics and cloud.

Sorbonne University (SU) is a leading French public research institution, renowned for its excellence in science and engineering. Its Faculty of Science and Engineering hosts over 3,500 researchers across top-tier laboratories. The Institute of Intelligent Systems and Robotics (ISIR) is a joint research unit comprising approximately 70 permanent researchers, and supported by a set of experimental robotics facilities, including nine distinct platforms.

The position

Dexterous manipulation is a problem at the intersection of control, perception, materials, and learning, and is characterized by challenges linked to long-horizon planning and memory; generalization to novel objects and tasks; perception under occlusion and uncertainty; high-dimensional control and coordination; and accurate contact dynamics and friction modeling. This internship will focus on the perception aspects of the task: robustly inferring cues about the 3D structure of the scene under occlusion and learning policies that effectively handle the sim-to-real gap, and exploiting these cues to solve complex manipulation tasks through machine learning.

To this end, we will exploit the latest 3D foundation models, which have been developed at NAVER LABS Europe, in particular the *St3R family [1-6], a set of breakthrough approaches in 3D reconstruction which, contrary to most existing methods, don’t require any camera parameters to perform direct 3D reconstruction from image content. Trained large-scale on massive amounts of real and simulated data, they work on any number of images, even when they do not overlap, offering a major simplification over traditional methods and significant versatility for diverse 3D vision challenges.

At NAVER LABS Europe, we also develop robust policies for agents in the real world for manipulation [7-8] and navigation [9-10] trained in simulation or on real data using reinforcement learning or imitation learning. A strong focus is put on spatial AI and on correctly modelling dynamics, and we have proposed latent geometric memory for robotics [11]. At SU, we develop model-based planners that enable sample-efficient policy learning [12] and introduce the specific adaptations required for their deployment on real robotic platforms [13].

The goal of this internship is to bring these two worlds together by investigating different strategies to enrich manipulation policies with the 3D awareness of *St3R models without explicitly reconstructing the scene. The core focus will be on leveraging latent 3D representations to enhance manipulation performance. A key objective is to evaluate the performance gain of these 3D environmental representations over their 2D counterparts. Furthermore, we will explore different methods for fusing this 3D representation with robot proprioception to achieve more robust control.

About the research team

In the Action group, we develop AI-driven decision-making capabilities that enable embodied agents to safely execute complex tasks in dynamic environments. To achieve autonomy in real-world everyday spaces, robots must learn from their interactions, understand how to best execute tasks specified by non-expert users, and do so in a safe and reliable manner. This requires sequential decision-making skills that integrate machine learning, adaptive planning, and control in uncertain environments, as well as the ability to solve hard combinatorial optimization problems. Our research combines expertise in reinforcement learning, computer vision, robotic control, sim-to-real transfer, large multimodal foundation models, and neural combinatorial optimization to design AI-based architectures and algorithms that enhance robot autonomy and robustness when completing complex tasks in constantly changing environments.

What we're looking for

• a candidate enrolled in an MSc program, aiming to pursue a PhD after the internship

• knowledge in computer vision with a deep learning background

• strong programming skills (even in cases where support from an LLM is not available)

Ideally starting in February or March 2026

Duration: six months

What we offer

• We foster a collaborative environment dedicated to ambitious, multidisciplinary projects that translate advanced research into impactful, real-world solutions, supported by 30+ years of experience in AI and related fields.
• Flexible work/life balance.

• We are an equal opportunity employer that hires based on skills, experience, and merit. We foster an inclusive and diverse workplace where all qualified candidates are considered fairly, regardless of background.

• We’re based in Meylan, close to Grenoble, a city that offers the perfect balance of urban life, cutting-edge research and technology, and spectacular mountain landscapes that provide countless opportunities to relax, recharge, and enjoy the outdoors.

All applications will be carefully considered, even if not all required skills are met. We value diverse backgrounds and the potential of each candidate, and we offer training to support the development of necessary skills.

NAVER LABS, co-located in Korea and France, is the organization dedicated to preparing NAVER’s future. Scientists at NAVER LABS Europe are empowered to pursue long-term research problems that, if successful, can have significant impact and transform NAVER. We take our ideas as far as research can to create the best technology of its kind. Active participation in the academic community and collaborations with world-class public research groups are, among others, important ways to achieve these goals. Teamwork, focus and persistence are important values for us.

 When applying for this position online, please don't forget to upload your CV and cover letter. Incomplete applications will not be considered.

NAVER LABS Europe is subject to French jurisdiction requiring organisations to stipulate that a job/internship is open to both women and men. None of our jobs/internships are gender specific.

References

1. Wang et al, DUSt3R: Geometric 3D Vision Made Easy, CVPR 2024.
2. Leroy et al., Grounding Image Matching in 3D with MASt3R, ECCV 2024.
3. Duisterhof et al., MASt3R-sfm: a fully-integrated solution for unconstrained structure-from-motion, 3DV 2024.
4. Cabon et al, MUSt3R: Multi-view Network for Stereo 3D Reconstruction, CVPR 2025.
5. Jang et al., Pow3r: Empowering unconstrained 3d reconstruction with camera and scene priors, CVPR 2025.
6. Zust et al, PanSt3R: Multi-view consistent panoptic segmentation, ICCV 2025.
7. Emukpere et al., Disentangled Object-Centric Image Representation for Robotic Manipulation, IROS 2025.
8. Kim et al., From exploration to control: learning object manipulation skills through novelty search and local adaptation, RSS 2021.
9. Bono et al., Learning to navigate efficiently and precisely in real environments, CVPR 2024. 
10. Janny et al., Reasoning in visual navigation of end-to-end trained agents: a dynamical systems approach, CVPR 2025. 
11. Sariyildiz et al., Kinaema: a recurrent sequence model for memory and pose in motion, NeurIPS 2025. 
12. El Asri et al., Physics-Informed Model and Hybrid Planning for Efficient Dyna-Style Reinforcement Learning. RLC 2024.
13. El Asri et al., RT-HCP: Dealing with Inference Delays and Sample Efficiency to Learn Directly on Robotic Platforms.

Réf: 9dfd9bff-80c0-4261-9638-c21335bb846a