Opportunities

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.

Download this offer

Thesis topic: Learning Generative World Models of Physical Dynamics

Context:

AI4Science is an emerging research field that investigates the potential of AI methods to advance scientific discovery, particularly through the modeling of complex natural phenomena. This fast- growing area holds the promise of transforming how research is conducted across a broad range of scientific domains. One especially promising application is in modeling complex dynamical systems that arise in fields such as climate science, earth science, biology, and fluid dynamics. A diversity of approaches is currently being developed, but this remains an emerging field with numerous open research challenges in both machine learning and domain-specific modeling.

This PhD project aims to investigate the next generation of AI models for physical dynamics. The objective is to develop generative world models that learn structured representations of physical systems and can efficiently model, predict, and reason about their evolution. The research will focus on applications such as fluid mechanics and climate science while addressing fundamental questions at the intersection of machine learning and scientific computing.

Research Directions:

The main objective of this PhD is to develop generative world models for physical dynamics that combine scalability, uncertainty modeling, and scientific consistency.

The research will explore several complementary directions:

– Learning transferable representations of physical dynamics, by developing latent representations that capture the underlying structure of physical systems and can generalize across multiple physical regimes and downstream tasks.

– Generative modeling of physical trajectories, using recent approaches such as diffusion models, flow matching, and stochastic interpolants to represent uncertainty, multimodality, and long- term evolution of complex dynamical systems.

– Physically consistent generative models, by integrating physical constraints and scientific priors into generative learning in order to produce solutions that remain both accurate and scientifically valid.

The exact research direction will be adapted to the candidate’s interests and background and may emphasize either methodological developments or applications to scientific domains such as fluid dynamics and climate modeling.

Position and Working Environment

The PhD studentship is a three years position starting in October/ November 2026. 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.

General information:
– Supervisor: Patrick Gallinari
– Collaboration as part of the PhD thesis: INRIA Paris, Institut d’Alembert Sorbonne Université
– Start date: November/ December 2026
– Note: The research topic is open and depending on the candidate profile could be oriented more on the theory or on the application side
– Host laboratory: ISIR (Institute of Intelligent Systems and Robotics), Pierre and Marie Curie Campus, 4 Place Jussieu, 75005 Paris.
– Keywords: AI4Science, deep learning, physics-aware deep learning, world models, generative models, foundation models

Application:
– Contact person: Patrick Gallinari
– Email: 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: 20/12/2026

Download this thesis offer

Thesis topic: Efficient Interaction through Information Maximization

Context:

Designing user interfaces is an iterative process. A designer creates an initial version based on their intuition and experience, then tests it with users and refines it based on feedback. This cycle is repeated until the interface is considered satisfactory. While effective, this approach is often slow and resource-intensive, as it requires many rounds of trial and error between designers and users.

To address this, researchers have proposed computational methods that automatically generate or optimize interface designs [2]. These methods rely on defining an objective (or “cost function”) that captures what makes a good interface, and then optimizing the design accordingly. However, a major limitation is that such objectives must be carefully crafted for each specific problem, which can be difficult.

This thesis explores a different approach: instead of designing task-specific objectives, it investigates whether general principles from information theory, specifically measures of how much information is exchanged between user and system, can be used to guide interface design in a more universal way.

Goal of the thesis: 

One promising idea in this direction is Bayesian Information Gain (BIG), introduced by Liu et al [1]. BIG is an interaction technique in which the system actively guides the user toward actions that are most informative (ie., actions that help the system quickly understand the user’s goal). Intuitively, the system tries to “ask the best possible question” at each step through its interface, so that the user’s response reveals as much useful information as possible. Although promising, BIG has several limitations. It is computationally expensive, has mainly been tested in small and discrete settings, and requires a calibration phase before use, which limits its practicality and adaptability. It also does not account for important usability aspects, such as proximality (the interface should not change too abruptly between steps, to avoid disorienting users) and intrinsic state value (some interface states may be inherently preferable, regardless of how informative they are). More broadly, BIG is only one of several recent approaches that use information-theoretic measures to design interactions (see positioning section). The first goal of this thesis is therefore to review and compare these different information-based approaches. This includes analyzing their theoretical properties, but also their practical characteristics, such as computational cost, ability to handle continuous or real-time interaction, and suitability for different application contexts.

The second goal is to address the limitations of BIG. Since many of these approaches rely on similar mathematical concepts (notably mutual information), improvements developed for BIG are likely to extend to other methods as well. Key challenges include estimating information efficiently from limited user data, and incorporating usability constraints such as smooth transitions and meaningful interface states. As part of this work, the candidate will develop a software library implementing both existing and newly proposed methods. The third goal is to evaluate these approaches in realistic applications (map navigation, text entry etc.)

Required Profile:

The candidate will have an interest and demonstrated expertise in computational modeling or machine learning. Interest and experience in experimental research and software programming, as well as knowledge of basic information theoretic notions will be appreciated but are not required.

General information:
– Supervisor: Julien Gori (CNRS)
– Possible co-supervisor: Olivier Rioul (Télécom Paris, Institut Polytechnique de Paris)
– Host laboratory: ISIR (Institute of Intelligent Systems and Robotics), Pierre and Marie Curie Campus, 4 Place Jussieu, 75005 Paris.

Application:
– Contact person: Julien Gori
– Email: gori@isir.upmc.fr
– Please send a CV, M1/M2 transcripts, and a copy of your master’s thesis (if available) when applying.
– Application deadline: rolling basis.

Download this thesis offer

Subject: Estimation and control of continuum robots with multiple sections

Project description

The past decade has seen the emergence of a new type of robot: continuum robots [Rus et al. 2015; Burgner-Kahrs et al. 2015; Gilbert et al. 2016; Morales Bieze et al. 2020; Childs et al. 2023; Tummers et al. 2025]. Their advantages over traditional rigid robots are numerous: a larger workspace, increased dexterity, greater compliance, the ability to navigate cluttered or tortuous environments, strong potential for miniaturization, the possibility of designing new kinematic architectures, as well as a diversification of applications, particularly in minimally invasive surgery or for handling fragile objects. The concept of continuum robots has thus enabled the emergence of a new paradigm in the field of dexterous robotics.

However, these major technological innovations still pose many challenges in terms of integration, perception, modeling, and control of these mechanisms. Consequently, many scientific issues—often fundamental—remain open before achieving effective integration of continuum robots into industrial and medical applications.

Recent advances in continuum robot control have enabled the implementation of closed-loop end-effector control strategies [Campisano et al. 2021], leading to submillimetric accuracy. However, implementing this type of control, particularly in minimally invasive surgical applications, requires robust and reliable estimation of the robot’s state.

Accurate state estimation of continuum robots is a complex problem and one of the missing building blocks for their widespread adoption. Among the various perception modalities in robotics, vision is by far the most commonly used. However, in continuum robotics, cameras are often unreliable—either due to the nature of the application, which prevents consistent visual contact with the robot (e.g., navigation in constrained environments, minimally invasive surgery), or due to self-occlusions caused by the robot’s own deformations.

This necessitates the exploration of alternative perception strategies. These generally rely on integrating additional sensors within a robotic architecture that is already highly constrained in terms of compactness. The resulting problem can then be divided into two major challenges:

– Maximizing the amount of information provided by small sensors about the robot’s state, without altering its intrinsic properties or compromising its ability to perform the intended tasks.

– Accurately estimating the robot’s state from partial or indirect measurements. This includes, for example, reconstructing the full state of the robot from position measurements at a limited number of points along its structure, or from force and torque measurements at its base. Due to the complexity of the deformations these robots undergo, these problems may be ill-posed and admit multiple solutions, particularly for the estimation of external loads.

Profile and skills required:

– Master graduation

– Knowledge in robotics, control, estimation and vision

– Skills in advanced programming

– Required level of English: Upper-intermediate. You can use the language effectively and express yourself accurately.

General information:

– Laboratory: Institute of Intelligent Systems and Robotics (ISIR), UMR 7222

– Affiliated institution: Sorbonne University

– Thesis title: Estimation and control of continuum robots with multiple sections

– Thesis supervisors: Faïz Ben Amar (Professor, Sorbonne University), Azad Artinian (Lecturer, Sorbonne University)

– Contact email: artinian@isir.upmc.fr

– Affiliation to a programme: ED SMAER

– Funding: Funding is subject to approval by the Doctoral School

Application procedure:

Candidates may apply online directly via ADUM until 18 May 2026.

To ensure a fair assessment, applications must include the documents listed below.

Documents to be included in the application:

– A detailed CV of the candidate and a cover letter explaining how their personal project aligns with that of the laboratory, on the chosen topic.

– A reference from the head of the candidate’s original Master’s programme, including ECTS credits earned, academic background, modules passed, ranking and the candidate’s average mark. As the final Master’s examination board will not have met by the application deadline, please include results from theoretical examinations and ECTS credits earned. The candidate’s ranking in the various semesters of their studies, and in particular in the first semester of the M2, are mandatory documents for a fair assessment of the candidate.

– The placement assessment form or the placement supervisor’s report

If your application is successful, you will be invited to an interview before a panel during which you will present your topic using 4 transparencies (further details will be provided at a later date).