Job offers
Context:
The LOGIE AI project aims to revolutionize logistics robotics through generative artificial intelligence. Coordinated by Enchanted Tools in collaboration with Hugging Face, NXP, and INRIA, and supported by ISIR (Sorbonne University), the project embraces an open-source approach to develop and integrate advanced language models into robots.
By embedding large language models (LLMs) and vision-language models (VLMs) into robotic systems, the project seeks to enable robots to understand, interpret, and respond to commands in a context-aware and precise manner. This integration will improve the efficiency of operations in complex environments such as logistics and human-robot interaction.
Missions:
The recruited candidate (post-doc or engineer) will contribute to one or several of the following tasks:
– Data collection and processing: designing and executing protocols to capture motion, interaction, and manipulation data, including simulation-based augmentation.
– Modeling and optimization: adapting and optimizing LLMs/VLMs for robotic applications, with a focus on real-time interaction, robustness, and efficiency.
– Embedded and hybrid solutions: deploying models on NXP processors and GPU clusters, evaluating hybrid local/cloud execution pipelines, and optimizing for energy efficiency.
– Validation and integration: running experiments on robotic platforms (e.g., Mirokai, Franka, Tiago, …) in both controlled and logistics-like environments.
– Contribution to open-source development: supporting collaborative development with project partners and the research community.
Post-doctoral researcher: expected to contribute to novel research directions, scientific publications, and supervision of junior students and interns.
Engineer: expected to contribute to system implementation, integration, experimental validation, and technical support.
Required profile:
– PhD in Robotics, Artificial Intelligence, Computer Science, or related field (required for post- doc).
– Master’s degree or engineering degree in Robotics, AI, Computer Science, or related disciplines (sufficient for the engineer position).
– Strong interest in AI for robotics, particularly in areas such as LLMs, VLMs, human-robot interaction, or robotic manipulation.
– Ability to work in a collaborative, interdisciplinary, and open-source environment.
Required skills:
– Proficiency in Python and common deep learning frameworks (e.g., PyTorch, TensorFlow).
– Experience with robotics middleware (ROS/ROS2) and robotic platforms.
– Knowledge in machine learning, natural language processing, or computer vision.
– Familiarity with embedded systems, GPU computing, or hybrid cloud/local architectures is a plus.
– Good communication skills in English (French is a plus, but not required).
This is a full-time position starting soon, based at ISIR, Sorbonne University, Paris. The selected candidate will join a vibrant research environment with access to advanced robotic platforms and international collaborations.
General information:
– Position type: Post-Doc or Engineer
– Contract start date: December 1, 2025
– Contract duration: 1 year, with the possibility of extension up to 2 years.
– Workload: 100%
– Desired experience: 1-4 years
– Desired level of education: Doctorate (post-doc) or Master’s degree/engineering school (engineer).
– Remuneration: Remuneration will be in accordance with the university’s current salary scale, depending on status (post-doctoral researcher or engineer).
– Host laboratory: ISIR (Institute for Intelligent Systems and Robotics), Pierre and Marie Curie Campus, 4 place Jussieu, 75005 Paris.
Contact:
– mahdi.khoramshahi@isir.upmc.fr
– Apply by email. Send your application by email, with [LogiE-IA] in the subject line, along with your resume and cover letter.
– Application deadline: October 15, 2025
Job title: Machine learning for human-robot collaboration
Context and objectives :
This position focuses on developing machine learning techniques to enhance human awareness in human-robot interaction by integrating situation assessment and action planning. The successful candidate will contribute to cutting-edge research at the intersection of robotics, artificial intelligence, and human interaction, with an emphasis on designing and evaluating robotic systems that facilitate seamless collaboration with humans.
The position is for 18 months contract, but there is a possibility to be extended depending on the performance and circumstances. The position is open at both the engineer and post-doctoral levels for candidates with a strong background in machine learning, human-machine interaction, or robotics.
Responsibilities:
– Develop advanced situation assessment techniques using machine learning to accurately represent user preferences, behaviors, and characteristics based multimodal data to efficiently plan actions.
– Collaborate with interdisciplinary teams including computer scientists, humanities, and designers to ensure the usability and effectiveness of developed techniques.
– Publish research findings in top-tier conferences and journals in the field of Human-Machine Interaction and Machine Learning (mainly at the post-doc level)
Requirements :
The successful candidate should have:
– Experience in human-machine interaction
– Good knowledge of Machine Learning Techniques
– Good knowledge of experimental design and statistics
– Excellent publication record
– Strong skills in Python
– Willing to work in multi-disciplinary and international teams
– Good communication skills
Additional information:
– Contract start date: as soon as possible
– Contract duration: 18 months
– Quota of work : 100%
– Level of education required: Master’s degree / Engineering school – Doctorate
– Salary: depending on experience
– Host laboratory: ISIR (Institut des Systèmes Intelligents et de Robotique), Campus Pierre et Marie Curie, 4 place Jussieu, 75005 Paris.
Application :
Interested candidates should submit the following by email in a single PDF file to: mohamed.chetouani[@]sorbonne-universite.fr with the subject: Application ML for Human-Robot Collaboration
– Curriculum vitae with 2 references (recommendation letters are also welcome)
– One-page summary of research background and interests
– At least three papers (either published, accepted for publication, or pre-prints) demonstrating expertise in one or more of the areas mentioned above
– Doctoral dissertation abstract and the expected date of graduation for a post-doc position levale (for those who are currently pursuing a Ph.D)
Application’s deadline: April 21, 2025.
Job title: 1.5-year post-doctorate in the ethics of open-ended learning in robotics and AI
Context:
The PILLAR-Robots European project (https://pillar-robots.eu) aims at developing a new generation of robots endowed with a higher level of autonomy, that are able to determine their own goals and establish their own strategies, creatively building on the experience acquired during their lifetime to fulfil the requests of their human designers/users in real-life application use-cases. To do so, it aims to implement a new theoretical extension of the reinforcement learning formalism, called the purpose framework*, which constitutes a domain-independent abstraction of a set of goals that expresses what the designer/user wants from the robot. This enables us to constraint robots’ autonomous exploration and learning to (1) focus on task- relevant objects and parts of the environment, and to (2) avoid reaching particular states of the environment or acting in ways that misalign with humans’ objectives, preferences and values.
*Baldassarre, G., Duro, R. J., Cartoni, E., Khamassi, M., Romero, A., & Santucci, V. G. (2024). Purpose for Open-Ended Learning Robots: A Computational Taxonomy, Definition, and Operationalisation. arXiv preprint arXiv:2403.02514.
Location and environment:
The post-doc position will be located in the Institute of Intelligent Systems and Robotics (ISIR, http://www.isir.upmc.fr), Paris, France. ISIR belongs to Sorbonne Université, CNRS and INSERM, and is located in the center of Paris, thus at walking distance from Seine River, from other academic institutions (La Sorbonne, Collège de France, Muséum d’Histoire Naturelle, Ecole Normale Supérieure, Université Paris Cité, Hôpital la Pitié Salpêtrière), and from famous monuments (Notre Dame, Conciergerie, Panthéon, Théâtre du Châtelet, Institut du Monde Arabe). Speaking or understanding French is not required. This work will be done in close collaborations with philosophers, engineers and computational neuroscientists of the PILLAR-Robots European consortium.
Missions:
The post-doc work will investigate possibilities of alignment and misalignement in such purposeful open-ended learning robots. In particular, research shall be done on how the combination of PILLAR’s robot motivational engine relying on the purpose framework with large language models (LLMs) can facilitate alignment, and under which conditions. Explainability, trust and efficient human-robot coordination will be evaluated in a set of real-world scenarios, corresponding to the project’s three use cases: a warehouse scenario, an agricultural robotics scenario and an edutainment scenario. An ethical analysis will be carried out by consulting consortium members on all three scenarios.
Required profile:
We are looking for highly motivated candidates with a strong academic record. An excellent background is expected in machine learning, AI, cognitive robotics or computational neuroscience. Significant experience in robot cognitive architectures, AI alignment, LLMs, or computational modeling for neuroscience or psychology will be appreciated. Prior knowledge in philosophy of mind and moral philosophy will be a plus. Eligibility: PhD degree in a quantitative discipline. There is no nationality or age criteria.
Required skills:
Mastery of reinforcement learning and game theory, very good level in applied maths, and advanced programming skills in modern C++ and python are required. Very good level of English (written, spoken).
General information :
– Position type : Post-doctorate
– Contract start date: 01/10/2025
– Contract duration: 18 months (until 30/03/2027)
– Working hours: 100%.
– Desired experience: from Beginner
– Level of education required : PhD
– Salary : Standard post-doc salary
– Source of funding : European PILLAR-Robots project
– Host laboratory: ISIR (Institut des Systèmes Intelligents et de Robotique), Campus Pierre et Marie Curie, 4 place Jussieu, 75005 Paris.
Contact person:
– Mehdi Khamassi
– Tel:+33650764492 Email : mehdi.khamassi@sorbonne-universite.fr
– Application: Send your application by e-mail, with [PILLAR post-doc application] in the subject line, a CV, a cover letter (max 2 pages) and a list of two references.
– Application deadline: 07/05/2025
Post-doc “Flexible robotics and digital twins for cardiac surgery”
Context:
This postdoctoral position is part of the RHU-ICELAND project, involving several academic, hospital, and industrial partners. The objective of the project is to develop a new transfemoral mitral valve annuloplasty solution that integrates intracardiac ultrasound imaging and robotics. This approach allows intervention on a beating heart without extracorporeal circulation, offering a mitral valve repair solution for high-risk patients who are ineligible for open-heart surgery and, in the long term, for most patients requiring mitral valve repair.
Direct annuloplasty involves affixing a ring or band directly onto the mitral annulus using anchors under echocardiographic and fluoroscopic guidance. The advantage of this technique is that it constrains the shape of the mitral annulus, closely replicating surgical mitral annuloplasty. The RHU-ICELAND project focuses on two key phases: developing a numerical model of the anatomy and the robotic system used to apply staples to the mitral valve, followed by designing and evaluating the robotic system, which is validated through numerical modelling.
Scientific Objectives:
Initially, the recruited postdoc will focus on the numerical modelling of anatomical structures (veins, heart, mitral valve, etc.). Preliminary work has already been carried out to design a numerical model of the heart and mitral valve with opening and closing cycles. The aim is to enhance this model for greater realism. The model will be used for clinician training, preoperative intervention planning, and validating the geometric, kinematic, and dynamic models of the robotic system (active catheter) during navigation from the entry point (femoral vein) to the target site (facing the mitral valve). The other medium- and long-term goal is to develop a realistic and, above all, patient-specific numerical model, meaning constructing the numerical model based on the patient’s preoperative images.
The recruited candidate will work closely with academic and clinical teams involved in the project, particularly when integrating the research into the final demonstrator. The postdoc will benefit from a stimulating research environment and access to clinical data provided by the project’s clinical and industrial partners. They will also participate in project management (meetings, decision-making, report writing, etc.).
Host Institution:
The recruited candidate will join the Institute of Intelligent Systems and Robotics (ISIR) at Sorbonne University and CNRS (Paris). ISIR is organized into several multidisciplinary teams, including RPI-Bio. Research areas include microrobotics, drones, surgical robotics, bionic prosthetics, social robots, and various intelligent and interactive systems (physical, virtual, or mixed-reality), as well as artificial intelligence. Applications address major societal challenges: health, the industry of the future, transportation, and personal services.
The RPI-Bio team (robotics, perception, and interaction for biomedical applications), to which the postdoc will be attached, conducts research in healthcare robotics on topics such as interactive systems for expert guidance (surgery), perception (visual and haptic), human-machine interfaces, telemedicine, and microrobotics. Recently labelled by Inserm, RPI-Bio has extensive experience in developing advanced robotic solutions for interventional medicine (orthopaedics, neurosurgery, ENT surgery, endovascular interventions, etc.).
Profile Sought:
– Expertise in robotics, mechatronics, simulation, and/or numerical modelling
– Advanced programming skills (C++, MATLAB, Python)
– Proficiency in a numerical simulation library for soft robots (e.g., SOFA) is a plus
– Enthusiasm for interdisciplinary research and a collaborative spirit
General information :
– Supervisors: Lingxiao Xun; Brahim Tamadazte
– Contract start date: as soon as possible
– Contract duration: 12 months, renewable for a further 12 months
– Salary: depending on experience
– Host laboratory: ISIR (Institut des Systèmes Intelligents et de Robotique), Campus Pierre et Marie Curie, 4 place Jussieu, 75005 Paris
Contact person:
Send a single PDF file containing: a CV, a cover letter, and any scientific articles you deem relevant to the application to lingxiao.xun@sorbonne-universite.fr and brahim.tamadazte@cnrs.fr. Please include ‘post-doc rhu’ in the subject line of the email.
Post-doc : Learning in robotics, with application to grasping
Context:
During the FET Proactive DREAM project (http://dream.isir.upmc.fr/) has been defined an approach for adaptive robotics based on open-ended learning. The main goal is to allow a robot to learn without requiring a careful preparation by an expert. This approach raises many challenges, notably learning with sparse reward, representation learning (for states and actions), model learning and exploitation, transfer learning, meta-learning and generalization. These topics are considered in simulation, but also on real robotics setup, notably in the context of grasping.
Missions:
This position aims at contributing to these topics in the context of several European projects, in particular SoftManBot, Corsmal, INDEX and Learn2Grasp. Calling upon previous works in the research team, the proposed approaches need to be easy to adapt to different robotic platforms and will thus be applied to different robots (Panda arm from Franka-Emika, Baxter, PR2 or TIAGO, for instance).
Required profile:
Candidates for the position must have a PhD degree in machine learning or related field in which robotics applications (either simulated or real) have been considered.
Required skills:
An excellent background is expected in machine learning as well as an experience in robotics. Excellent programming skills in Python are expected.
General Information:
- Position Type: Post-doctoral researcher
- Contract duration: 24 months
- Level of education required: PhD
- Remuneration : Remuneration according to experience
- Location: ISIR (Institut des Systèmes Intelligents et de Robotique), Campus Pierre et Marie Curie, 4 place Jussieu, 75005 Paris.
Contact person:
- Stephane Doncieux
- stephane.doncieux(at)sorbonne-universite.fr
- Send your application by email, with a CV and a cover letter.
PhD offers
Thesis Title: Deep Generative Models of Physical Dynamics
Context
AI4Science is an emerging field investigating the potential of AI to advance scientific discovery, with deep learning playing a central role in modeling complex natural phenomena. Within this context, deep generative modeling—which already enables the synthesis of high-dimensional data across modalities such as text, images, and audio—is now opening new avenues for simulating and understanding complex physical systems.
This PhD project aims to explore and advance generative deep learning architectures—such as diffusion models, flow-matching networks, and autoregressive transformers—for modeling complex physical dynamical systems arising in domains such as climate, biology, and fluid mechanics. These models hold strong potential for learning flexible, data-driven representations of physical laws. By developing generalizable, cross-physics generative models, this research contributes to the broader vision of AI4Science: accelerating scientific discovery through learned simulation and abstraction.
Research Directions
The overarching research question is: Can we develop generative models that learn structured, physically grounded representations of dynamical systems—enabling synthesis, adaptation, and generalization across physical regimes and multiphysics settings? It unfolds into several complementary directions:
Latent Generative Models for Physical Dynamics
The objective is to design generative models—such as diffusion, flow-matching, or autoregressive models—that learn compact and interpretable latent representations of spatiotemporal dynamics governed by PDEs. These models should:
– Capture uncertainty and multimodality in solution trajectories.
– Generalize across parametric variations.
Learning Across Multiphysics Systems
To enable transfer learning across heterogeneous physics, we will explore shared latent representations across families of PDEs:
– Using encode–process–decode frameworks.
– Applying contrastive or multitask training to uncover reusable physical abstractions.
– Designing models invariant to space/time resolution and units.
This direction builds toward foundation-like models that capture generalizable physics priors across simulation families.
Few-Shot and In-Context Generalization to New Physics
To support scientific modeling in data-scarce settings, we will develop methods for few-shot generalization such as:
– Fine-tuning latent priors to new PDE systems using limited examples.
– Exploring meta-learning and prompt-based adaptation techniques (inspired by in- context learning in language models).
– Incorporating known physical constraints into the generative process.
The goal is to enable rapid and physically consistent adaptation to previously unseen dynamics with minimal data and supervision.
Position and Working Environment
The PhD studentship is a three years position starting in October/November 2025. It does not include teaching obligation, but it is possible to engage if desired. The PhD candidate will work at Sorbonne Université (S.U.), in the center of Paris. He/She will integrate the MLIA team (Machine Learning and Deep Learning for Information Access) at ISIR (Institut des Systèmes Intelligents et de Robotique).
Required Profile:
Master degree in computer science or applied mathematics, Engineering school. Background and experience in machine learning. Good technical skills in programming.
Additional information :
– Thesis Supervisor: Patrick Gallinari, patrick.gallinari@sorbonne-universite.fr
– Possible Co-Supervisor: Nicolas Baskiotis, nicolas.baskiotis@sorbonne-universite.fr
– Thesis Collaboration: INRIA St Etienne and Lille, Institut d’Alembert Sorbonne University, CNAM Paris
– Host Laboratory: ISIR (Institute of Intelligent Systems and Robotics), Pierre and Marie Curie Campus, 4 Place Jussieu, 75005 Paris.
– Start date: November/ December 2025
– Note: The research topic is open and depending on the candidate profile could be oriented more on the theory or on the application side
– Keywords: AI4Science, deep learning, physics-aware deep learning, generative models, foundation models
Contact :
– Patrick Gallinari : patrick.gallinari@sorbonne-universite.fr
– Please send a cv, motivation letter, grades obtained in master, recommendation letters when possible to patrick.gallinari@sorbonne-universite.fr
– Application deadline: 20/12/2025
Thesis title: Frugal reinforcement learning for the control of the Atalante X exoskeleton
Context:
The use of exoskeletons to assist walking in individuals with disabilities is a promising solution for improving their quality of life. However, for these devices to be effective and accepted by users, it is crucial that they are personalized to each individual’s needs and capabilities. In this context, reinforcement learning can play an important role by enabling the exoskeleton control to adapt to the specific characteristics of each user, with the goal of achieving smooth, responsive movements that do not interfere with the user’s own actions. Indeed, reinforcement learning allows the controlled system to gradually refine its decisions through interaction with the environment, adapt to unforeseen usage conditions, and improve over time. However, its practical effectiveness may be limited by the need to collect a large amount of data before obtaining a high-performing control policy.
Scientific Objective and Project Description:
The objective of this PhD thesis will be to study and implement a frugal approach based on robotic priors that increase learning efficiency and enable rapid personalization of the system’s control policy — ideally within just a few hours of use.
These robotic priors refer to a set of predefined biases derived from expert knowledge of the robotic control task at hand. They will help guide the agent’s exploration of the action space and reduce the number of interactions required to obtain an effective and personalized control policy. At least two types of priors will be considered as starting points for this work:
– The first type of prior is temporal, based on the sequential nature of movements. For example, in the case of exoskeletons assisting gait, movements can be decomposed into several predefined phases that are synchronized with the user’s walking pattern. Knowing these phases helps guide the learning process toward acceptable solutions and partially addresses the credit assignment problem (i.e. the difficulty of attributing delayed rewards to the correct actions).
This direction will build upon previous work [1] using a “divide and conquer” approach, which has led to significantly improved learning efficiency over the state of the art in simulated biped locomotion tasks from single demonstrations.
– The second type of prior is geometric, based on knowledge of the three-dimensional workspace and certain key measurements. For instance, using an inertial measurement unit (IMU), one can obtain accurate and responsive data about orientation relative to the vertical axis — a crucial variable in balance control. By combining this information with other sensory and system characteristics (proprioception, dimensions, forward and inverse kinematics), the goal will be to structure and weight the agent’s observations and define auxiliary reward signals that progressively guide learning toward successful task completion. A curriculum will be defined — that is, a sequence of subtasks of increasing difficulty — to incrementally adjust the control policy parameters. For example, the curriculum may begin with static balance, proceed to executing a single step with the exoskeleton, and then to chaining multiple steps, ensuring at each stage that performance on earlier subtasks is not degraded.
These approaches will accelerate online learning, but the starting point will always be a policy pretrained in simulation, using a reasonably accurate dynamic model of the exoskeleton. This preliminary training will follow the well-known domain randomization approach and produce an initial control policy, which will then be refined through online reinforcement learning. To safely implement this online learning process, precise kinematic and dynamic constraints will be defined to ensure the safety of all experiments.
Finally, while the proposed approach will be specifically designed for controlling exoskeletons for gait assistance, it should also aim to be generalizable, at least to some extent. To verify that the method is not overly tailored to a single system, we will consider testing it on other devices that are simpler than a full walking exoskeleton.
[1] Chenu, A., Serris, O., Sigaud, O., & Perrin-Gilbert, N. (2022). Leveraging sequentiality in reinforcement learning from a single demonstration. arXiv preprint arXiv:2211.04786.
Required Profile:
We are looking for a candidate with a Master’s degree (Master 2) and solid experience in Python development, along with good knowledge of deep reinforcement learning. A strong interest in experimentation on real robotic systems is essential, and hands-on experience in this area is a major asset. The ideal candidate will be autonomous, persevering, and rigorous, with a strong interest in engineering and practical work.
Required skills:
Technical Skills
– Python Development: Advanced proficiency in Python with well-structured code practices (OOP, modularity, testing, version control).
– Deep Reinforcement Learning (Deep RL): Familiarity with standard algorithms (DQN, PPO, SAC…), their theoretical foundations, and practical implementation.
– Machine Learning Frameworks: Experience with libraries such as PyTorch, TensorFlow, or JAX.
– Robotic System Handling (ideally): Experience in controlling real robotic platforms (robotic arms, mobile robots, etc.).
– RL and Robotic Simulation: Knowledge of tools such as Mujoco and Gymnasium.
Soft Skills
– Autonomy: Ability to work independently, identify challenges, and proactively seek solutions.
– Perseverance: Enthusiasm for tackling complex technical problems and patience in the face of long, sometimes unstable, real-world robotic experiments.
– Engineering Mindset: Interest in prototyping, debugging (both hardware and software), and building robust systems.
– Communication: Ability to document and share results and methodologies effectively.
General Information:
– PhD supervisor: Nicolas Perrin-Gilbert
– The thesis is scheduled to start in September or October 2025.
– Laboratory: ISIR (Institut des Systèmes Intelligents et de Robotique), Campus Pierre et Marie Curie, 4 place Jussieu, 75005 Paris.
Contact:
– Nicolas Perrin-Gilbert ; perrin@isir.upmc.fr
– Send your application by e-mail, with [Frugal reinforcement learning for the control of the Atalante X exoskeleton] in the subject line, a CV and a covering letter.
– Application deadline : June 6, 2025
Thesis topic: Generation of 3D models of anatomical structures by deep learning from 2D images
Context:
The number of spinal surgery operations is rising steadily, both for trauma-related cases and for degenerative pathologies. In the latter case, the increase can be explained by the ageing of the population and possibly by changes in lifestyle, such as an increase in sedentary lifestyles and obesity. In addition to this increase in the number of cases of spinal degeneration, there has also been an increase in diagnostic capabilities, new treatment modalities and a broadening of clinical indications. As an example, around 450,000 arthrodeses were performed in 2014 in the United States alone, an increase of 118% since 1998. Lumbar fusions alone accounted for around 200,000 operations in the United States in 2015, 62% more than in 2004, at a cost of several tens of thousands of dollars per operation and a total annual cost of several billion dollars. Worldwide, several million pedicle screws are used in spinal surgery every year.
Scoliosis is treated surgically. It involves inserting screws into the vertebral pedicles to stabilise the spine. These screws are connected by rods to correct the abnormal curvature in three dimensions. This technique ensures rigid fixation, promoting vertebral fusion and improving functional and aesthetic results. The placement of these screws is complex and difficult, particularly in certain patients. The literature reports that, on average, without a support system, around 24% of pedicle screws are incorrectly positioned, ultimately leading to potentially serious neurological symptoms.
In parallel with an increase in the volume of pedicle screws placed, the field of computer-assisted surgery has seen a major boom, with the development of three-dimensional navigation solutions, the use of robotics and augmented/mixed reality methods. Each of these solutions has its own advantages and disadvantages. For example, navigation systems make pedicle screw placement more accessible to junior surgeons and reduce certain risks and complications. However, they are expensive, can extend the duration of surgical procedures, and are less accurate when not used optimally. Robotics and augmented/mixed reality provide added clinical value, particularly in the phase of inserting screws into the pedicles, given that the insertion trajectories are known.
Scientific objectives:
In a recent collaboration between Sorbonne University, Trousseau Hospital and the company SpineGuard, we removed several scientific barriers to the safe insertion of pedicle screws. In particular, we have developed a functional robotic platform for spinal surgery and real-time breach detection methods enabling us to stop the screw insertion task before a breach is created in the spinal canal. However, the approaches developed to date assume that the 3D trajectory of screw insertion into the pedicle is known precisely in the real world.
The aim of this doctoral project is precisely to address the problem of estimating the 3D trajectory of screw insertion during surgery. Currently, trajectories are planned by the surgeon on the basis of pre-operative imaging (e.g. CT scan). These trajectories are then transposed onto the patient on the day of the operation, assuming that the pre-operative 3D model corresponds exactly to what the surgeon sees.
In addition, the surgical procedure itself induces deformations in real time, which compromise the registration and planning of trajectories. In order to improve this registration and therefore the accuracy of the surgical procedure, we would like to explore the use of a conventional non-irradiating 2D visual sensor (e.g. RGB camera) to access the upper (visible) part of the spine in real time during the operation. In other words, this means observing the spinous processes and the transverse processes. The scientific challenge we propose to explore is to reconstruct each vertebra in 3D using 2D visual information from RGB images using deep learning methods. This relatively recent discipline has received growing interest from the computer vision community, significantly advancing the state of the art. Until recently, the reconstruction of a 3D model required several 2D views of the object/scene or, at the very least, a large database of annotated 2D image – 3D model (e.g. CAD) pairs, which severely limited its deployment in disciplines such as surgery due to the lack of annotated data.
To meet the scientific challenges and clinical needs, several avenues of work are envisaged. The first involves generating synthetic 2D images from 3D models, followed by domain adaptation so as to benefit from self-supervised learning of the 2D/3D transformation. This method, proposed for rigid objects, could be extended to non-rigid objects using displacement vector fields. Another promising approach is the result of recent work enabling a 3D model to be reconstructed from a simple 2D view. The tools used by these methods, and in particular triplanes, could be used for our application, by deploying them locally for each vertebra. A graph-based scene context network could then refine the initial 3D pose of each vertebra and their relative arrangement.
We already have a database of segmented and annotated vertebrae from patient scans, which we are continually adding to. As a result, the prediction of the 3D shape of vertebrae from 2D images can be enriched by the 3D information from the scans.
Support team:
It will consist of Brahim Tamadazte (DR CNRS, ISIR, SU), Catherine Achard (PU, ISIR, SU) and Raphaël Vialle (PUPH, APHP, Hôpital Trousseau, SU). Collaboration between ISIR and Hôpital Trousseau is well established, notably through participation in the EU H2020 FAROS project, the FHU SpineMed2 project, and the ANR RODEO project, resulting in several co-supervision of theses (3) and trainee doctors (2). These collaborations have enabled us to structure a significant research activity around the use of robotics, computer vision and AI to improve spinal surgery protocols.
In this type of arrangement between ISIR and a hospital centre, we usually involve one or more interns who come to do 6 to 12 month Masters placements. This enables greater interaction between the medical world and academic research. This model has amply demonstrated its translational capacity, leading to the creation of several companies (Basecamp Vascular, MovaLife) or the marketing of medical devices via specialist companies (Endocontrol, Koelis, GEMA, Moon Surgical, GE Healthcare, Robeauté and SpinGuard).
Application:
There are two funding possibilities (IUIS and ED SMAER) for this thesis project. Double applications are therefore required (deadline 5 May 2025):
– ED SMAER: https://adum.fr/as/ed/voirproposition.pl?langue=&site=edsmaer&matricule_prop=62877
– IUIS: https://lime3-app3.sorbonne-universite.fr/index.php/283197?lang=fr
Contacts :
– Catherine Achard, PU Sorbonne University (ISIR): catherine.achard@sorbonne-universite.fr
– Brahim Tamadazte, DR CNRS (ISIR): brahim.tamadazte@cnrs.fr
Internship offers
Subject: Intelligent Cursor Control from Discrete Inputs for users with limited input capabilities
Abstract:
Interacting with low dimension/bandwidth inputs (eye blinks, EEG signals, simple keypads) is crucial for assistive technologies. For example, people with severe motor impairments may be unable to move a computer mouse.
An example is KeyNav (https://github.com/jordansissel/keynav and https://www.youtube.com/watch?v=Uot3Cs9YwOA ), which uses four directional keys to move a cursor by recursively splitting the screen space. While simple, this dichotomic approach is not optimal in terms of speed and efficiency.
Internship Objectives:
The goal of this internship is to develop a smarter cursor control system using Bayesian Information Gain (BIG)[1,2], a method that works by Bayesian updating and maximizing the information-theoretic concept of mutual information. Instead of splitting space uniformly, the BIG algorithm positions the cursor where it maximally reduces uncertainty about the user’s intended target. BIG requires target information (where the potential cursor can point towards) to be operationalized.
Approach:
* TargetFinder [3], a computer vision tool that detects visible on-screen targets which we recently developped, will be used to access target information in a cross-platform way, just using pixels.
* TargetFinder output will be integrated with a BIG-based decision module that positions the cursor.
* Implement a prototype allowing cursor control with only four keys (up, down, left, right).
* Evaluate the system experimentally against KeyNav.
Expected Outcome:
A working prototype demonstrating efficient, information-driven cursor movement without a mouse, potentially improving accessibility and control for users with limited input capabilities.
Required Profile: Computer/Data science, intelligent systems with an interest in human computer interaction
Required skills: programming (mainly Python), basic mastery of probability / Bayesian inference. Optionnally : Computer vision, empirical studies, pyQT
General information:
– Supervisor: Julien Gori
– Start date of internship: between January and March 2026
– Duration of internship: 5 to 6 months
– Required level of education: Master’s degree or final year of engineering school
– Host laboratory: ISIR (Institute for Intelligent Systems and Robotics), Pierre and Marie Curie Campus, 4 place Jussieu, 75005 Paris.
Contact and application:
– Julien Gori; Email: gori@isir.upmc.fr
– Send your application by email, with [internship subject] in the subject line, a CV and cover letter, and a copy of your transcripts.
Subject: Teleoperation control law development, implementation and testing for robot hand- arm system
Abstract:
Dexterous manipulation is a challenging problem in robotics because it involves strong interaction with the environment, and this environment is not always known. Several studies have shown the importance of tactile feedback to improve robots’ manipulation capabilities.
This internship proposes to contribute to the development and validation of a bidirectional teleoperation interface for a dexterous robotic hand mounted on a 7 DOF arm. The system consists of a leader (arm + hand directly controlled by the human operator and transmitting the forces experienced by the follower back to the human operator) and a follower (arm + hand copying the movements of the leader system and interacting with the environment).
The work will focus on developing the control law followed by both systems. This control law must simultaneously allow the follower to replicate the leader’s movements and enable the leader to transmit the forces experienced by the follower back to the operator. This control law will then be tested and validated.
Detailed plan:
– Creation of a ROS 2 environment that enables communication with two arms and two robotic hands on a central PC. (w1 – 4)
– Development of the control law (Justification of architectural choices based on literature review). (w5 – 8)
– Implementation of the control law in Python (or C++ if necessary for performance reasons) (w9 – 12)
– Testing and optimization of parameters, then characterization of system behavior (w13 – 16)
– Collection of a dataset for a task and training of a policy (w17 – 24)
Internship Objectives:
The main objective is to develop and validate a bidirectional teleoperation interface for a dexterous robotic hand, allowing a human operator to remotely control a robot while feeling the forces it experiences. The work involves creating a control law that ensures movement synchronization between the leader system (controlled by the human) and the follower system (which interacts with the environment), while transmitting haptic feedback. The final goal is to collect data to train a machine learning policy based on this teleoperation interface.
Required Profile: Masters student in robotics engineering
Required skills: ROS 2, Python, C++, Control systems and Machine learning
General information:
– Supervisor: Mahdi Khoramshahi
– Start date of internship: 1 December 2025
– Duration of internship: 6 months
– Required level of education: Master’s degree
– Host laboratory: ISIR (Institute for Intelligent Systems and Robotics), Pierre and Marie Curie Campus, 4 place Jussieu, 75005 Paris.
Contact:
– Mahdi Khoramshahi; mahdi.khoramshahi@isir.upmc.fr
– Send your application by email, with [internship subject] in the subject line, a CV and a cover letter.
– Application deadline: 15 October 2025
