2025Activity report​​​‌Project-TeamLARSEN

Scientific​​ context

So far, only​​​‌ a few autonomous robots‌ have been deployed for‌​‌ a long time (weeks,​​ months, or years) outside​​​‌ of factories and laboratories.‌ They are mostly mobile‌​‌ robots that simply “move​​ around” (e.g., vacuum cleaners​​​‌ or museum “guides”) and‌ data collecting robots (e.g.,‌​‌ boats or underwater “gliders”​​ that collect data about​​​‌ the water of the‌ ocean).

A large part‌​‌ of the long-term autonomy​​ community is focused on​​​‌ simultaneous localization and mapping‌ (SLAM), with a recent‌​‌ emphasis on changing and​​ outdoor environments 21,​​​‌ 28. A more‌ recent theme is life-long‌​‌ learning: during long-term deployment,​​ we cannot hope to​​​‌ equip robots with everything‌ they need to know,‌​‌ therefore some things will​​ have to be learned​​​‌ along the way. Most‌ of the work on‌​‌ this topic leverages machine​​ learning and/or evolutionary algorithms​​​‌ to improve the ability‌ of robots to react‌​‌ to unforeseen changes 21​​​‌, 26.

Main​ challenges

The first major​‌ challenge is to endow​​ robots with a stable​​​‌ situation awareness in open​ and dynamic environments. This​‌ covers both the state​​ estimation of the robot​​​‌ by itself as well​ as the perception/representation of​‌ the environment. Both problems​​ have been claimed to​​​‌ be solved but it​ is only the case​‌ for static environments 25​​.

In the Larsen​​​‌ team, we aim at​ deployment in environments shared​‌ with humans which imply​​ dynamic objects that degrade​​​‌ both the mapping and​ localization of a robot,​‌ especially in cluttered spaces.​​ Moreover, when robots stay​​​‌ longer in the environment​ than for the acquisition​‌ of a snapshot map,​​ they have to face​​​‌ structural changes, such as​ the displacement of a​‌ piece of furniture or​​ the opening or closing​​​‌ of a door. The​ current approach is to​‌ simply update an implicitly​​ static map with all​​​‌ observations but without attempt​ at distinguishing the suitable​‌ changes. For localization in​​ not-too-cluttered or not-too-empty environments,​​​‌ this is generally sufficient​ since a significant fraction​‌ of the environment should​​ remain stable. But for​​​‌ life-long autonomy, and in​ particular for navigation, the​‌ quality of the map,​​ and especially the knowledge​​​‌ of the stable parts,​ is primordial.

A second​‌ major obstacle to moving​​ robots outside of labs​​​‌ and factories is their​ fragility: Current robots​‌ often break in a​​ few hours, if not​​​‌ a few minutes. This​ fragility mainly stems from​‌ the overall complexity of​​ robotic systems, which involve​​​‌ many actuators, many sensors,​ and complex decisions, and​‌ from the diversity of​​ situations that robots can​​​‌ encounter. Low-cost robots exacerbate​ this issue because they​‌ can be broken in​​ many ways (high-quality material​​​‌ is expensive), because they​ have low self-sensing abilities​‌ (sensors are expensive and​​ increase the overall complexity),​​​‌ and because they are​ typically targeted towards non-controlled​‌ environments (e.g., houses rather​​ than factories, in which​​​‌ robots are protected from​ most unexpected events). More​‌ generally, this fragility is​​ a symptom of the​​​‌ lack of adaptive abilities​ in current robots.

Angle​‌ of attack

To solve​​ the state estimation problem,​​​‌ our approach is to​ combine classical estimation filters​‌ (Extended Kalman Filters, Unscented​​ Kalman Filters, or particle​​​‌ filters) with a Bayesian​ reasoning model in order​‌ to internally simulate various​​ configurations of the robot​​​‌ in its environment. This​ should allow for adaptive​‌ estimation that can be​​ used as one aspect​​​‌ of long-term adaptation. To​ handle dynamic and structural​‌ changes in an environment,​​ we aim at assessing,​​​‌ for each piece of​ observation, whether it is​‌ static or not.

We​​ also plan to address​​​‌ active sensing to improve​ the situation awareness of​‌ robots. Literally, active sensing​​ is the ability of​​​‌ an interacting agent to​ act so as to​‌ control what it senses​​ from its environment with​​​‌ the typical objective of​ acquiring information about this​‌ environment. A formalism for​​ representing and solving active​​​‌ sensing problems has already​ been proposed by members​‌ of the team 20​​ and we aim to​​ use it to formalize​​​‌ decision-making problems for improving‌ situation awareness.

Situation awareness‌​‌ of robots can also​​ be tackled by cooperation,​​​‌ whether it be between‌ robots or between robots‌​‌ and sensors in the​​ environment (deployed in sensorized​​​‌ environments) or between robots‌ and humans. This is‌​‌ in rupture with classical​​ robotics, in which robots​​​‌ are conceived as self-contained.‌ But, in order to‌​‌ cope with as diverse​​ environments as possible, these​​​‌ classical robots use precise,‌ expensive, and specialized sensors,‌​‌ whose cost prohibits their​​ use in large-scale deployments​​​‌ for service or assistance‌ applications. Furthermore, when all‌​‌ sensors are on the​​ robot, they share the​​​‌ same point of view‌ on the environment, which‌​‌ is a limit for​​ perception. Therefore, we propose​​​‌ to complement a cheaper‌ robot with sensors distributed‌​‌ in a target environment.​​

To address the fragility​​​‌ problem, the traditional approach‌ is to first diagnose‌​‌ the situation, then use​​ a planning algorithm to​​​‌ create/select a contingency plan.‌ But, again, this calls‌​‌ for both expensive sensors​​ on the robot for​​​‌ the diagnosis and extensive‌ work to predict and‌​‌ plan for all the​​ possible faults that, in​​​‌ an open and dynamic‌ environment, are almost infinite.‌​‌ An alternative approach is​​ then to skip the​​​‌ diagnosis and let the‌ robot discover by trial‌​‌ and error a behavior​​ that works in spite​​​‌ of the damage with‌ a reinforcement learning algorithm‌​‌ 33, 26.​​ However, current reinforcement learning​​​‌ algorithms require hundreds of‌ trials/episodes to learn a‌​‌ single, often simplified, task​​ 26, which makes​​​‌ them impossible to use‌ for real robots and‌​‌ more ambitious tasks. We​​ therefore need to design​​​‌ new trial-and-error algorithms that‌ will allow robots to‌​‌ learn with a much​​ smaller number of trials​​​‌ (typically, a dozen). We‌ think the key idea‌​‌ is to guide online​​ learning on the physical​​​‌ robot with dynamic simulations.‌ For instance, in our‌​‌ recent work, we successfully​​ mixed evolutionary search in​​​‌ simulation, physical tests on‌ the robot, and machine‌​‌ learning to allow a​​ robot to recover from​​​‌ physical damage 27,‌ 1.

A key‌​‌ functionality of autonomy is​​ the capacity to make​​​‌ decision. Our approach is‌ to address it within‌​‌ the framework of sequential​​ decision making which can​​​‌ be studied using Markov‌ Decision Processes and other‌​‌ derived models. A stronger​​ research direction of the​​​‌ team consists in their‌ theoretical study, which involves‌​‌ stochastic (or sometimes deterministic)​​ games.

A final approach​​​‌ to address fragility is‌ to deploy several robots‌​‌ or a swarm of​​ robots or to make​​​‌ robots evolve in an‌ active environment. We will‌​‌ consider several paradigms such​​ as (1) those inspired​​​‌ from collective natural phenomena‌ in which the environment‌​‌ plays an active role​​ for coordinating the activity​​​‌ of a huge number‌ of biological entities such‌​‌ as ants and (2)​​ those based on online​​​‌ learning 24. We‌ envision to transfer our‌​‌ knowledge of such phenomenon​​ to engineer new artificial​​​‌ devices such as an‌ intelligent floor (which is‌​‌ in fact a spatially​​​‌ distributed network in which​ each node can sense,​‌ compute and communicate with​​ contiguous nodes and can​​​‌ interact with moving entities​ on top of it)​‌ in order to assist​​ people and robots (see​​​‌ the principle in 31​, 24, 19​‌).

Scientific​​​‌ context

Interaction with the​ environment is a primordial​‌ requirement for an autonomous​​ robot. When the environment​​​‌ is sensorized, the interaction​ can include localizing, tracking,​‌ and recognizing the behavior​​ of robots and humans.​​​‌ One specific issue lies​ in the lack of​‌ predictive models for human​​ behavior and a critical​​​‌ constraint arises from the​ incomplete knowledge of the​‌ environment and the other​​ agents.

On the other​​​‌ hand, when working in​ the proximity of or​‌ directly with humans, robots​​ must be capable of​​​‌ safely interacting with them,​ which calls upon a​‌ mixture of physical and​​ social skills. Currently, robot​​​‌ operators are usually trained​ and specialized but potential​‌ end-users of robots for​​ service or personal assistance​​​‌ are not skilled robotics​ experts, which means that​‌ the robot needs to​​ be accepted as reliable,​​​‌ trustworthy and efficient 36​. Most Human-Robot Interaction​‌ (HRI) studies focus on​​ verbal communication 32 but​​​‌ applications such as assistance​ robotics require a deeper​‌ knowledge of the intertwined​​ exchange of social and​​​‌ physical signals to provide​ suitable robot controllers.

Main​‌ challenges

We are here​​ interested in building the​​​‌ bricks for a situated​ HRI addressing both the​‌ physical and social dimension​​ of the close interaction,​​​‌ and the cognitive aspects​ related to the analysis​‌ and interpretation of human​​ movement and activity.

The​​​‌ combination of physical and​ social signals into robot​‌ control is a crucial​​ investigation for assistance robots​​​‌ 34 and robotic co-workers​ 30. A major​‌ obstacle is the control​​ of physical interaction (precisely,​​​‌ the control of contact​ forces) between the robot​‌ and the human while​​ both partners are moving.​​​‌ In mobile robots, this​ problem is usually addressed​‌ by planning the robot​​ movement taking into account​​​‌ the human as an​ obstacle or as a​‌ target, then delegating the​​ execution of this “high-level”​​​‌ motion to whole-body controllers,​ where a mixture of​‌ weighted tasks is used​​ to account for the​​​‌ robot balance, constraints, and​ desired end-effector trajectories 18​‌.

The first challenge​​ is to make these​​​‌ controllers easier to deploy​ in real robotics systems​‌, as currently they​​ require a lot of​​​‌ tuning and can become​ very complex to handle​‌ the interaction with unknown​​ dynamical systems such as​​​‌ humans. Here, the key​ is to combine machine​‌ learning techniques with such​​ controllers.

The second challenge​​​‌ is to make the​ robot react and adapt​‌ online to the human​​ feedback, exploiting the​​​‌ whole set of measurable​ verbal and non-verbal signals​‌ that humans naturally produce​​ during a physical or​​​‌ social interaction. Technically, this​ means finding the optimal​‌ policy that adapts the​​ robot controllers online, taking​​​‌ into account feedback from​ the human. Here, we​‌ need to carefully identify​​ the significant feedback signals​​ or some metrics of​​​‌ human feedback. In real-world‌ conditions (i.e., outside the‌​‌ research laboratory environment) the​​ set of signals is​​​‌ technologically limited by the‌ robot's and environmental sensors‌​‌ and the onboard processing​​ capabilities.

The third challenge​​​‌ is for a robot‌ to be able to‌​‌ identify and track people​​ on board. The​​​‌ motivation is to be‌ able to estimate online‌​‌ either the position, the​​ posture, or even moods​​​‌ and intentions of persons‌ surrounding the robot. The‌​‌ main challenge is to​​ be able to do​​​‌ that online, in real-time‌ and in cluttered environments.‌​‌

Angle of attack

Our​​ key idea is to​​​‌ exploit the physical and‌ social signals produced by‌​‌ the human during the​​ interaction with the robot​​​‌ and the environment in‌ controlled conditions, to learn‌​‌ simple models of human​​ behavior and consequently to​​​‌ use these models to‌ optimize the robot movements‌​‌ and actions. In a​​ first phase, we will​​​‌ exploit human physical signals‌ (e.g., posture and force‌​‌ measurements) to identify the​​ elementary posture tasks during​​​‌ balance and physical interaction.‌ The identified model will‌​‌ be used to optimize​​ the robot whole-body control​​​‌ as prior knowledge to‌ improve both the robot‌​‌ balance and the control​​ of the interaction forces.​​​‌ Technically, we will combine‌ weighted and prioritized controllers‌​‌ with stochastic optimization techniques.​​ To adapt online the​​​‌ control of physical interaction‌ and make it possible‌​‌ with human partners that​​ are not robotics experts,​​​‌ we will exploit verbal‌ and non-verbal signals (e.g.,‌​‌ gaze, touch, prosody). The​​ idea here is to​​​‌ estimate online from these‌ signals the human intent‌​‌ along with some inter-individual​​ factors that the robot​​​‌ can exploit to adapt‌ its behavior, maximizing the‌​‌ engagement and acceptability during​​ the interaction.

Another promising​​​‌ approach already investigated in‌ the Larsen team is‌​‌ the capability for a​​ robot and/or an intelligent​​​‌ space to localize humans‌ in its surrounding environment‌​‌ and to understand their​​ activities. This is an​​​‌ important issue to handle‌ both for safe and‌​‌ efficient human-robot interaction.

Simultaneous​​ Tracking and Activity Recognition​​​‌ (STAR) 35 is an‌ approach we want to‌​‌ develop. The activity of​​ a person is highly​​​‌ correlated with his position,‌ and this approach aims‌​‌ at combining tracking and​​ activity recognition to make​​​‌ one benefit from the‌ other. By tracking the‌​‌ individual, the system may​​ help infer its possible​​​‌ activity, while by estimating‌ the activity of the‌​‌ individual, the system may​​ make a better prediction​​​‌ of his/her possible future‌ positions (especially in the‌​‌ case of occlusions). This​​ direction has been tested​​​‌ with simulator and particle‌ filters 23, and‌​‌ one promising direction would​​ be to couple STAR​​​‌ with decision making formalisms‌ like partially observable Markov‌​‌ decision processes (POMDPs). This​​ would allow us to​​​‌ formalize problems such as‌ deciding which action to‌​‌ take given an estimate​​ of the human location​​​‌ and activity. This could‌ also formalize other problems‌​‌ linked to the active​​ sensing direction of the​​​‌ team: how should the‌ robotic system choose its‌​‌ actions in order to​​​‌ better estimate the human​ location and activity (for​‌ instance by moving in​​ the environment or by​​​‌ changing the orientation of​ its cameras)?

Another issue​‌ we want to address​​ is robotic human body​​​‌ pose estimation. Human body​ pose estimation consists of​‌ tracking body parts by​​ analyzing a sequence of​​​‌ input images from single​ or multiple cameras.

Human​‌ posture analysis is of​​ high value for human​​​‌ robot interaction and activity​ recognition. However, even though​‌ the arrival of new​​ sensors like RGB-D cameras​​​‌ has simplified the problem,​ it still poses a​‌ great challenge, especially if​​ we want to do​​​‌ it online, on a​ robot and in realistic​‌ world conditions (cluttered environment).​​ This is even more​​​‌ difficult for a robot​ to bring together different​‌ capabilities both at the​​ perception and navigation level​​​‌ 22. This will​ be tackled through different​‌ techniques, going from Bayesian​​ state estimation (particle filtering),​​​‌ to learning, active and​ distributed sensing.

5.1.1 pepper_driver_ext

• Name:‌​‌
  Additional ROS drivers for​​ the Pepper robot.
• Keywords:​​​‌
  Pepper robot, Driver
• Functional‌ Description:

  These drivers extend‌​‌ the robot functionalities accessible​​ in ROS in complement​​​‌ to the original drivers.‌

  Namely they add:

  • handling‌​‌ of the autonomous state​​ of the robot,
  • (un)loading​​​‌ and (de)activating dialog topics,‌
  • text-to-speech functionality,
  • ALMemory access,‌​‌
  • publication of detected landmarks,​​
  • tablet handling,
  • LED control,​​​‌
  • fixed /cmd_vel driver.

  They‌ are designed for Ubuntu‌​‌ 16.04 with ROS kinetic​​ and Python 2.7 to​​​‌ work alongside the official‌ ROS drivers. For newer‌​‌ platforms (Ubuntu 20.04 and​​ later), tools and Dockerfiles​​​‌ are available to use‌ the Pepper and its‌​‌ drivers from ROS 2.​​

• Release Contributions:
  Lower resource​​​‌ consumption via on-demand subscription.‌
• URL:
  https://­colas.­gitlabpages.­inria.­fr/­pepper_driver_ext/­index.­html
• Contact:
  Francis‌​‌ Colas
• Participants:
  Francis Colas,​​ Vincent Colotte

5.1.2 PACR​​​‌ Project

• Keywords:
  Robotics, Simulation,‌ Teaching
• Functional Description:

  Set‌​‌ of tools and documentation​​ to implement and learn​​​‌ a complex robotic navigation‌ task. The task requires‌​‌ probabilistic planning, path planning,​​ and path following with​​​‌ obstacle avoidance. The software‌ proposes a simulated warehouse‌​‌ environment with several robots​​ and libraries and code​​​‌ to accomplish the project.‌

  This is planned as‌​‌ a 3x2h project lab​​ work for the PACR​​​‌ module on computer science‌ M2 IA²VR at Université‌​‌ de Lorraine.

• Release Contributions:​​
  Support for:

  • Ubuntu 22.04​​​‌ with ROS 2 Humble‌ and Gazebo Fortress
  • Ubuntu‌​‌ 24.04 with ROS 2​​ Jazzy and Gazebo Harmonic​​​‌
• URL:
  https://­ia2vr.­gitlabpages.­inria.­fr/­pacr/­index.­html
• Contact:
  Francis‌ Colas
• Participants:
  Francis Colas,‌​‌ Alexis Scheuer, Vincent Thomas​​
• Partner:
  Université de Lorraine​​​‌

5.1.3 ILIAR Project

• Name:‌
  Autonomous driving simulation and‌​‌ project for computer science​​ teaching
• Keywords:
  Robotics, Autonomous​​​‌ Cars, Teaching, Machine learning,‌ Computer vision
• Functional Description:‌​‌

  Set of tools and​​ documentation defining a simulation​​​‌ of an autonomous car‌ on a circuit so‌​‌ as to implement a​​ controller based on machine​​​‌ learning.

  The documentation is‌ a step-by-step tutorial to‌​‌ implement the following steps:​​

  • system discovery,
  • teleoperation,
  • data​​​‌ collection,
  • actual machine learning,‌
  • controller implementation,
  • controller evaluation.‌​‌

  This project is the​​ main part of the​​​‌ ILIAR module in computer‌ science M2 IA²VR at‌​‌ Université de Lorraine.

• Release​​ Contributions:
  Support for:

  • Ubuntu​​​‌ 22.04 with ROS 2‌ Humble and Gazebo Fortress.‌​‌
  • Ubuntu 24.04 with ROS​​ 2 Jazzy and Gazebo​​​‌ Harmonic.
• URL:
  https://­ia2vr.­gitlabpages.­inria.­fr/­iliar/­index.­html
• Contact:‌
  Francis Colas
• Participants:
  Francis‌​‌ Colas, Jérémy Fix
• Partner:​​​‌
  CentraleSupélec

A simple​​ random game model for​​​‌ a better analysis of​ deterministic game-solving algorithms

Participants:​‌ Raphael Boige, Amine​​ Boumaza, Bruno Scherrer​​​‌.

Deterministic game-solving algorithms​ are conventionally analyzed in​‌ the light of their​​ average-case complexity on some​​​‌ random model. We have​ introduced a new simple​‌ probabilistic model that incrementally​​ constructs game-trees using a​​​‌ fixed level-wise conditional distribution.​ By enforcing ancestor dependencies,​‌ a critical structural feature​​ of real-world games, our​​​‌ framework generates problems with​ adjustable difficulty while retaining​‌ some form of analytical​​ tractability. For several algorithms,​​​‌ including AlphaBeta and Scout,​ we have derived recursive​‌ formulas characterizing their average-case​​ complexities under this model.​​​‌ These allow us to​ rigorously compare algorithms on​‌ deep game-trees, where Monte-Carlo​​ simulations are no longer​​​‌ feasible. This work has​ been published in 6​‌

$\epsilon$-Optimally Solving Two-Player​​ Zero-Sum POSGs

Participants: Olivier​​​‌ Buffet.

Collaboration with​ Jilles Dibangoye, Erwan Escudie,​‌ and Matthia Sabatelli (University​​ of Groningen).

Many robotic​​​‌ scenarios involve multiple interacting​ agents, robots or humans,​‌ e.g., security robots in​​ public areas. After addressing​​​‌ in the past the​ collaborative setting, where all​‌ agents share one objective​​ 2, we have​​​‌ applied a similar approach​ in the important 2-player​‌ zero-sum setting, i.e., with​​ two competing agents, and​​​‌ proposed an algorithm for​ partially observable Stochastic Games​‌ (POSGs), turning the problem​​ into an occupancy Markov​​​‌ game, and deriving bounding​ approximators that build on​‌ two types of continuity​​ properties: Lipschitz-continuity, and convexity​​​‌ and concavity properties.

This​ year, we have introduced​‌ a lossless reduction from​​ zs-POSGs to transition-independent zs-SGs,​​​‌ enabling the principled application​ of a broad class​‌ of DP-based methods. We​​ show empirically that point-based​​​‌ value iteration (PBVI) algorithms,​ applied via this reduction,​‌ produce $\epsilon$-optimal strategies​​ across a range of​​​‌ benchmark domains, consistently matching​ or outperforming existing state-of-the-art​‌ methods. This work and​​ the obtained results have​​​‌ been published in 8​.

Partially Observable Monte-Carlo​‌ Graph Search

Participants: Yang​​ You, Olivier Buffet​​​‌, Vincent Thomas.​

This work also involves​‌ the Oxford Robotics Institute​​ and the UK Atomic​​​‌ Energy Authority where Yang​ You (former member of​‌ Larsen project-team) is pursuing​​ post-doctoral research.

Currently, large​​​‌ partially observable Markov decision​ processes (POMDPs) are often​‌ solved by sampling-based online​​ methods which interleave planning​​​‌ and execution phases. However,​ a pre-computed offline policy​‌ is more desirable in​​ POMDP applications with time​​​‌ or energy constraints. But​ previous offline algorithms are​‌ not able to scale​​ up to large POMDPs.​​​‌

We worked on a​ new sampling-based algorithm, the​‌ partially observable Monte-Carlo graph​​ search (POMCGS) to solve​​​‌ large POMDPs offline. Instead​ of developing a tree​‌ while performing Monte-Carlo simulations,​​ POMCGS folds this tree​​​‌ on the fly, thus​ generating a policy graph,​‌ reducing computations and increasing​​ the interpretability of the​​​‌ policy. By adding progressive​ widening and observation clustering,​‌ POMCSGs is able to​​ address some continuous POMDPs.​​​‌ This work and the​ obtained results on classical​‌ POMDP benchmarks have been​​ published in 12.​​

Post-Hoc Interpretation of POMDP​​​‌ Policies

Participants: Olivier Buffet‌.

Collaboration with Geoffrey‌​‌ Laforest, Alexandre Niveau, and​​ Bruno Zanuttini (GREYC, Université​​​‌ de Caen Normandie) as‌ part of the ANR‌​‌ project EpiRL.

Dynamic epistemic​​ logic allows reasoning about​​​‌ an agent's knowledge and‌ its evolution given the‌​‌ occurrence of events. Recent​​ work has developed epistemic​​​‌ planning, i.e., seeking for‌ a sequence of actions‌​‌ that leads to some​​ state of knowledge (e.g.,​​​‌ knowing the value of‌ some variable, or whether‌​‌ some fact is true​​ or not). The EpiRL​​​‌ ANR project aims at‌ performing a similar task‌​‌ through (possibly deep) reinforcement​​ learning, i.e., learning a​​​‌ behavior by trial and‌ error.

In this context,‌​‌ the PhD thesis of​​ Geoffrey Laforest in Caen​​​‌ (supervised by Bruno Zanuttini‌ and Alexandre Niveau) looks‌​‌ in particular at the​​ choice of suitable representations​​​‌ for the state of‌ knowledge (which should be‌​‌ compact, and ideally embed​​ probabilities, due to the​​​‌ stochastic dynamics). These representations‌ can then be used‌​‌ as input of control​​ policies.

We proposed to​​​‌ redescribe policies into mappings‌ defined on features of‌​‌ the current belief state,​​ built in a systematic​​​‌ manner from state features.‌ Such a mapping can‌​‌ in turn be represented​​ by an intelligible object,​​​‌ like a decision tree,‌ thereby providing an interpretable‌​‌ representation of the policy​​ as a whole. We​​​‌ moreover showed how our‌ approach allows to explain‌​‌ the decision taken by​​ an agent at each​​​‌ step of an interaction‌ with the environment. This‌​‌ provides an end-to-end process,​​ starting from a policy​​​‌ computed by any solver,‌ and ending with an‌​‌ explanation of each decision​​ made at execution time.​​​‌ In 14, 9‌, 15, we‌​‌ formally define our approach,​​ investigate related computational problems,​​​‌ and report on experiments‌ on several families of‌​‌ problems.

Task-planning for human robot‌ collaboration

Participants: Yang You‌​‌, Francis Colas,​​ Olivier Buffet, Vincent​​​‌ Thomas.

This work‌ has been done in‌​‌ part during the former​​ ANR project Flying CoWorker.​​​‌

This work focuses on‌ high-level decision making for‌​‌ collaborative robotics. When a​​ robot has to assist​​​‌ a human worker, it‌ has no direct access‌​‌ to the worker's current​​ intention or preferences but​​​‌ has to adapt its‌ behavior to help the‌​‌ human complete his task.​​

Human-robot collaboration often necessitates​​​‌ the robot to adapt‌ to the uncertainty of‌​‌ human objectives and their​​ induced behaviors. This may​​​‌ require the robot to‌ have a human model‌​‌ to anticipate human partners'​​ objectives and predict their​​​‌ actions, which is typically‌ learned by the robot‌​‌ through available human data.​​ However, in complex collaboration​​​‌ tasks, a chicken-and-egg problem‌ arises because human data‌​‌ cannot be collected without​​ a collaborative robot policy​​​‌ in the first place.‌ We had previously proposed‌​‌ to describe the human-robot​​ collaboration task with Markov​​​‌ decision models and to‌ solve the chicken-and-egg problem‌​‌ through a probabilistic planning​​ algorithm. This year, we​​​‌ have contributed an online‌ version of this approach.‌​‌ This online framework can​​​‌ automatically derive a human​ model without real human​‌ data and plan robust​​ robot actions to support​​​‌ human partners with respect​ to their uncertain objectives​‌ and behaviors. Through experiments​​ with a human-robot co-working​​​‌ scenario, we demonstrate that​ our online method outperforms​‌ the previous offline approach​​ in terms of scalability​​​‌ and the ability to​ plan robot actions within​‌ a bounded time. This​​ work has been presented​​​‌ in 11.

Explicability​ and interpretability in probabilistic​‌ planning

Participants: Salomé Lepers​​, Sophie Lemonnier,​​​‌ Olivier Buffet, Vincent​ Thomas.

Part of​‌ this work is a​​ collaboration with Shuwa Miura​​​‌ and Shlomo Zilberstein from​ University of Massachusetts (UMass)​‌ at Amherst.

In a​​ human-agent collaboration scenario, some​​​‌ properties of the agent​ behavior can be useful​‌ for the human and​​ sometimes allow a better​​​‌ collaboration. These properties include,​ for instance, legibility (legible​‌ behaviors convey intentions, i.e.,​​ actual task at hand,​​​‌ via action choices), explicability​ (explicable behaviors conform to​‌ observers' expectations, i.e., they​​ appear to have some​​​‌ purpose), and predictability (a​ behavior is usually considered​‌ predictable if it is​​ easy to guess the​​​‌ end of an on-going​ trajectory). Recent theoretical frameworks​‌ allow formalizing such properties​​ and proposing algorithms to​​​‌ enforce them. In Salomé​ Lepers' PhD thesis, we​‌ build in particular on​​ Miura and Zilberstein's OAMDP​​​‌ framework (observer aware Markov​ decision process), where an​‌ agent interacts with a​​ stochastic environment while trying​​​‌ to optimize a criterion​ that depends on an​‌ external observer's belief.

We​​ have first looked in​​​‌ particular at predictability, where​ the end of the​‌ current trajectory may highly​​ depend on the outcome​​​‌ of each action, and​ thus proposed that predictability​‌ should be about minimizing​​ the number of errors​​​‌ when an external observer​ is asked repeatedly to​‌ guess the next action​​ or next state. This​​​‌ has been formalized in​ a variant of the​‌ observer-aware Markov Decision Process​​ (OAMDP) formalism, naturally coming​​​‌ with simple algorithms that​ efficiently find optimal solutions.​‌ We conducted in silico​​ and in vivo experiments​​​‌ (where actual humans observe​ the behavior of artificial​‌ agents) on simple grid-world​​ problems to validate the​​​‌ approach, as presented in​ 5.

The main​‌ direction we have then​​ been pursuing as part​​​‌ of Salomé Lepers' PhD​ thesis is to allow​‌ for an observer with​​ partial and noisy observability​​​‌ (the agent knowing exactly​ what the observer perceives),​‌ which led to introducing​​ the PO-OAMDP formalism (​​​‌partially observable OAMDPs).​ We have shown that​‌ this formalism is a​​ strict generalization of OAMDPs,​​​‌ and that the current​ state of the system,​‌ along with the observer's​​ belief about that state,​​​‌ make for a sufficient​ statistic for optimal planning.​‌ This allowed proposing a​​ variant of the heuristic​​​‌ search value iteration algorithm​ that relies on pointwise​‌ and cone approximators, the​​ later leveraging Lipschitz-continuity. We​​​‌ also demonstrated that, in​ stochastic shortest-path problems, some​‌ information-oriented criteria may not​​ induce policies that reach​​​‌ a terminal state with​ probability 1, what can​‌ be fixed by adding​​ a per-step cost. These​​ results, with illustrations of​​​‌ the resulting behaviors on‌ various problems, are presented‌​‌ in 10, 16​​.

During this year,​​​‌ Salomé also wrote her‌ PhD thesis manuscript and‌​‌ defended her PhD thesis​​ on December 17, 2025.​​​‌

Adaptive control of collaborative‌ robots for preventing musculoskeletal‌​‌ disorders

Participants: Aya Yaacoub​​, Francis Colas,​​​‌ Vincent Thomas.

This‌ work is done in‌​‌ collaboration with Pauline Maurice​​ from the HUCEBOT team.​​​‌

The use of collaborative‌ robots in direct physical‌​‌ collaboration with humans constitutes​​ a possible answer to​​​‌ musculoskeletal disorders: not only‌ can they relieve the‌​‌ worker from heavy loads,​​ but they could also​​​‌ guide them towards more‌ ergonomic postures. In this‌​‌ context, one objective of​​ the ROOIBOS Project is​​​‌ to build adaptive robot‌ strategies that are optimal‌​‌ regarding productivity but also​​ the long-term health and​​​‌ comfort of the human‌ worker, by adapting the‌​‌ robot behavior to the​​ human's physiological state.

In​​​‌ previous works, we proposed‌ to use Partially Observable‌​‌ Markov Decision Processes (POMDP)​​ to compute a robot​​​‌ policy taking into account‌ the long-term consequences of‌​‌ the biomechanical demands on​​ the human worker's joints​​​‌ (joint loading) and to‌ distribute the efforts among‌​‌ the different joints during​​ the execution of a​​​‌ repetitive task. The proposed‌ platform merges within the‌​‌ same framework several works​​ conducted in the Larsen​​​‌ team, namely virtual human‌ modeling and simulation, fatigue‌​‌ estimate and decision making​​ in the face of​​​‌ uncertainties. We also proposed‌ an approach to automatically‌​‌ extract a small discrete​​ set of relevant actions​​​‌ from the continuous action‌ space by indentifying and‌​‌ gathering relevant actions through​​ short-term planning (greedy-based selection​​​‌ approach) phases.

During this‌ year, we designed an‌​‌ experiment to validate the​​ effectiveness of the proposed​​​‌ POMDP-based planning approach for‌ fatigue mitigation with human‌​‌ subjects, in a human-robot​​ co-manipulation repeated task. Although​​​‌ the protocol hes been‌ fully described, the experiments‌​‌ have not yet been​​ conducted because of time​​​‌ limitations.

During 2025, Aya‌ wrote her PhD thesis‌​‌ manuscript, and with the​​ approval of the reviewers,​​​‌ her thesis defense is‌ scheduled for February 3,‌​‌ 2026.

Identifying human movement​​ strategies for human-robot collaboration​​​‌

Participants: Vincent Thomas,‌ Francis Colas.

This‌​‌ work is done in​​ collaboration with former post-doc​​​‌ Jessica Colombel and with‌ Pauline Maurice from the‌​‌ HUCEBOT team.

In human-robot​​ physical collaboration, it is​​​‌ necessary that the robot‌ can anticipate the whole-body‌​‌ posture of the human​​ co-worker to enable a​​​‌ seamless and efficient collaboration.‌ When co-manipulating an object,‌​‌ the human posture is​​ partly guided by the​​​‌ pose of the robot‌ end-effector. However owing to‌​‌ the high kinematic redundancy​​ of the human body,​​​‌ an infinity of postures‌ can in theory be‌​‌ adopted for a same​​ hand pose. In practice,​​​‌ human movements are largely‌ stereotyped, which drastically reduces‌​‌ the number of observed​​ solutions. Yet some diversity​​​‌ remains, which we refer‌ to as “movement strategies”.‌​‌ The objective of this​​ work is to develop​​​‌ a methodology to identify‌ human movement strategies in‌​‌ a manual task, and​​​‌ explore the relations between​ movement strategies annd factors​‌ such as anthropometry and​​ physical fatigue. During the​​​‌ postdoctoral work of Jessica​ Colombel, we designed an​‌ experiment and conducted a​​ large data collection campaign,​​​‌ to acquire human motion​ data in a repetitive​‌ manual task to work​​ on. We started to​​​‌ analyze the data, and​ explored diffusion methods to​‌ cluster the data in​​ different movement strategies. Inverse​​​‌ optimal control is also​ a method that we​‌ plan to explore.

This​​ line of work led​​​‌ to a preliminary publication​ as an abstract and​‌ presentation in a French​​ biomechanics conference 4.​​​‌

PhD grant with Airbus​​​‌

Participants: Olivier Buffet,​ Aubin Delaveau.

Collaboration​‌ with Florent Teichtel-Königsbuch (Airbus).​​

The thesis is funded​​​‌ by Airbus to contribute​ to the development of​‌ new aircraft with improved​​ hybrid energy management, assuming​​​‌ that both kerosene and​ hydrogen fuel cells can​‌ be used to produce​​ electricity. This involves both​​​‌ high-level energy management for​ the whole duration of​‌ a flight, and low-level​​ (thus real-time) control of​​​‌ a physical (electric) system​ under various conditions.

8.1.1 ANR​​​‌ : EpiRL

Participants: Olivier​ Buffet.

• Program:
  ANR​‌
• Project title:
  Apprentissage par​​ renforcement épistémique
• Duration:
  February​​​‌ 2023 – February 2027​
• Coordinator:
  François Schwarzentruber (École​‌ normale supérieure de Lyon)​​
• Partner institutions:
  IRIT CNRS,​​​‌ DAVI, IRISA ENS de​ Rennes, GREYC Université de​‌ Caen-Normandie, ENS de Lyon​​
• Abstract:

  EpiRL project aims​​​‌ at investigating the combination​ of epistemic planning and​‌ reinforcement learning (RL), by​​ proposing new algorithms that​​​‌ are efficient, adaptive, and​ capable of computing decisions​‌ relying on theory of​​ knowledge and belief. We​​​‌ expect from this approach​ an efficiency in the​‌ generation of epistemic plans,​​ while decisions made by​​​‌ RL algorithms will be​ explainable. Moreover, the algorithms​‌ of EpiRL will be​​ tested and evaluated within​​​‌ a real application that​ exploits autonomous agents.

  The​‌ project will address the​​ weaknesses of both epistemic​​​‌ planning and RL: on​ the one hand, existing​‌ epistemic planning algorithms are​​ costly, do not adapt​​​‌ to the environment, and​ concepts are hand-crafted and​‌ are not learned; on​​ the other hand, in​​​‌ reinforcement learning, agents adapt​ to their environments but​‌ are unable to reason​​ about beliefs of other​​​‌ agents. The newly developed​ algorithms will combine the​‌ strengths of both fields.​​

  Four workpackages are proposed:​​​‌

  1. Study representations of states​
  2. Develop RL algorithms
  3. Study​‌ representations of policies
  4. Validating​​ the algorithms with the​​​‌ industrial partner DAVI, in​ particular, through the development​‌ of a debunking chatbot​​ whose use case will​​​‌ apply to raising awareness​ about environmental issues.

  In​‌ this project our responsibility​​ lies in the study​​​‌ and definition of representations​ for the knowledge state​‌ and the policy, for​​ reinforcement learning algorithms.

9.1.1 Scientific events:​‌ organisation

9.1.2 Scientific events: selection​​

9.1.3 Journal

9.1.4 Scientific​​ expertise

• Francis Colas was​​​‌ expert for an ASTRID‌ project.
• Francis Colas was‌​‌ expert for a CIFRE​​ project.

9.1.5 Research administration​​​‌

• Amine Boumaza is a‌ member of comité de‌​‌ centre.
• Olivier Buffet is​​ a member of the​​​‌ FSSSCT.
• Francis Colas is‌ member of the following‌​‌ commissions: Comipers, ComiDoc, Comité​​ de Centre.
• Vincent Thomas​​​‌ is a member of‌ the commission IES du‌​‌ centre.

9.2.1 Teaching

Vincent Thomas‌​‌ is co-responsible of the​​ parcours "Intelligence Artificielle et​​​‌ ses Applications en Vision‌ et Robotique" of “Master‌​‌ 2 Informatique”, Univ. Lorraine,​​ France.

• Master: Amine Boumaza​​​‌ , “Recherche Locale Stochastique‌ et Métaheuristiques”, 30h eq.‌​‌ TD, M1 Informatique, Univ.​​ Lorraine, France.
• Master: Francis​​​‌ Colas , “Planification de‌ trajectoires”, 17h eq. TD,‌​‌ M2 Informatique “Intelligence Artificielle​​ et ses Applications en​​​‌ Vision et Robotique”, Univ.‌ Lorraine, France.
• Master: Francis‌​‌ Colas , “Ingénierie Logicielle​​ pour l'Intelligence Artificielle et​​​‌ la Robotique”, 15h eq.‌ TD, M2 Informatique “Intelligence‌​‌ Artificielle et ses Applications​​ en Vision et Robotique”,​​​‌ Univ. Lorraine, France.
• Master:‌ Francis Colas , “Situation‌​‌ Intégratrice”, 36h eq. TD,​​ M2 Informatique “Intelligence Artificielle​​​‌ et ses Applications en‌ Vision et Robotique”, Univ.‌​‌ Lorraine, France.
• Master: Alexis​​ Scheuer , “Contrôle d'exécution”,​​​‌ 17h eq. TD, M2‌ Informatique “Intelligence Artificielle et‌​‌ ses Applications en Vision​​ et Robotique”, Univ. Lorraine,​​​‌ France.
• Master: Alexis Scheuer‌ , “Éléments de robotique”,‌​‌ 4h eq. TD, M2​​ MEEF, Univ. Lorraine, France.​​​‌
• Master: Alexis Scheuer ,‌ “Robotique autonome”, 30h eq.‌​‌ TD, M1 Informatique, Univ.​​ Lorraine, France.
• Master: Vincent​​​‌ Thomas , “Planification et‌ apprentissage”, 17h eq. TD,‌​‌ M2 Informatique “Intelligence Artificielle​​ et ses Applications en​​​‌ Vision et Robotique”, Univ.‌ Lorraine, France.
• Master: Vincent‌​‌ Thomas , “Game Design”,​​ 37h eq. TD, M1​​​‌ Sciences Cognitives, Univ. Lorraine,‌ France.
• Master: Vincent Thomas‌​‌ , “Agent Intelligent”, 25h​​ eq. TD, M1 Sciences​​​‌ Cognitives, Univ. Lorraine, France‌
• Bachelor: Alexis Scheuer ,‌​‌ “Optimisation”, 53h eq. TD,​​ L3 Informatique (FST), Univ.​​​‌ Lorraine, France.
• Bachelor: Alexis‌ Scheuer , “Conception”, 8h‌​‌ eq. TD, L3 Informatique​​ (FST), Univ. Lorraine, France.​​​‌
• Bachelor: Alexis Scheuer ,‌ “Introduction à la robotique”,‌​‌ 40h eq. TD, L2​​ Informatique (FST), Univ. Lorraine,​​​‌ France.
• Bachelor: Alexis Scheuer‌ , “Bureautique”, 40h eq.‌​‌ TD, L1 Informatique (FST),​​ Univ. Lorraine, France.
• Bachelor:​​​‌ Vincent Thomas , “Conception‌ et Programmation”, 168h eq.‌​‌ TD, BUT Informatique, Univ.​​ Lorraine, France.
• Bachelor: Vincent​​​‌ Thomas , ”Optimisation et‌ bases de l'apprentissage automatique”,‌​‌ 46h eq. TD, BUT​​​‌ Informatique, Univ. Lorraine, France.​

9.2.2 Supervision

• PhD defended:​‌ Salomé Lepers , “Interpretability​​ and Explicability in Probabilistic​​​‌ Planning”, started in October​ 2022, defended on 2025-12-17,​‌ Olivier Buffet (advisor), Vincent​​ Thomas (co-advisor).
• PhD in​​​‌ progress: Aya Yaacoub ,​ “User-specific planning of a​‌ collaborative robot behavior to​​ help prevent musculoskeletal disorders”,​​​‌ started in December 2021,​ Francis Colas (advisor), Pauline​‌ Maurice (co-advisor), Vincent Thomas​​ (co-advisor), ROOIBOS project.
• PhD​​​‌ in progress: Raphael Boige​ , “Tree search algorithms:​‌ beyond Monte Carlo tree​​ search”, started in November​​​‌ 2024, Bruno Scherrer (advisor),​ Amine Boumaza (co-advisor).
• PhD​‌ in progress: Aubin Delaveau​​ , “Safe and Robust​​​‌ Management of Energy in​ Hybrid Aircrafts”, started in​‌ February 2025, Olivier Buffet​​ (advisor), in collaboration with​​​‌ Florent Teichteil-Königsbuch (Airbus).
• PhD​ in progress: Antonin Rousseau​‌ , “Robot navigation under​​ exogeneous uncertainty”, started in​​​‌ November 2025, Francis Colas​ (advisor), Alexis Scheuer (co-advisor).​‌
• Master M2: Paul Loisil​​ , “Incertitude de mouvement”,​​​‌ March to August 2025,​ Francis Colas (co-advisor), Alexis​‌ Scheuer (co-advisor).
• Master M2:​​ Nicolas Queignec , “Planification​​​‌ de chemin en environnement​ inconnu et dynamique”, March​‌ to July 2025, Francis​​ Colas (advisor).
• Master M1:​​​‌ Emna Debbech , “Apprentissage​ profond pour la résolution​‌ de $\rho$-POMDP. Application​​ au problème de recherche​​​‌ de proie dans un​ labyrinthe.”, July to September​‌ 2025, Olivier Buffet (co-advisor),​​ Vincent Thomas (co-advisor) in​​​‌ collaboration with Guénaël Cabanes​ (co-advisor).
• Master M1: Nada​‌ El Hanafi , ”Infrastructure​​ de navigation pour l'exploration​​​‌ robotique”, June to August​ 2025, Francis Colas (advisor).​‌
• Bachelor (L3): Jarod Galbrun​​ , "Preuve formelle pour​​​‌ la planification probabiliste", June​ to July 2025, Olivier​‌ Buffet (co-advisor) in collaboration​​ with Pierre-Jean Spaenlehauer (EPI​​​‌ CARAMBA, co-advisor).
• Bachelor (BUT3):​ Adrien Naigeon , "Plateforme​‌ d’apprentissage pour la planification​​ automatique", April to June​​​‌ 2025, Vincent Thomas (co-advisor),​ Olivier Buffet (co-advisors).
• Bachelor​‌ (L2): Camille Desplas ,​​ “Algorithmes de recherche de​​​‌ plus courts chemins”, May​ to June 2025, Francis​‌ Colas (advisor).

9.2.3 Juries​​

• Olivier Buffet was reviewer​​​‌ for the PhD of​ Hector Kohler (Cristal/INRIA, Université​‌ de Lille) and for​​ the HDR of Guillaume​​​‌ Lozenguez (IMT Nord Europe,​ Université de Lille).
• Francis​‌ Colas was reviewer for​​ the PhD of Camille​​​‌ Charrier (Université Grenoble Alpes,​ LPNC).
• Bruno Scherrer was​‌ reviewer for the PhD​​ of Chiara Mignacco (Université​​​‌ Paris-Saclay).

9.2.4 Educational and​ pedagogical outreach

• Alexis Scheuer​‌ participated at the "fête​​ de la science 2025"​​​‌ by organizing and co-animating​ an introductory robotics workshop​‌ for pupils, Univ. Lorraine.​​
• Vincent Thomas was a​​​‌ member of the organizing​ comité of the "séminaire​‌ de pédagogie universitaire" (2025),​​ Univ. Lorraine.
• Vincent Thomas​​​‌ was a member of​ the organizing comité of​‌ the Game Jam for​​ university educators (2025), Univ.​​​‌ Lorraine.
• Vincent Thomas organized​ two workshops from journée​‌ ISN for secondary school​​ teachers, at Loria (Nancy).​​​‌

9.3.1 Specific​ official responsibilities in science​‌ outreach structures

• Amine Boumaza​​ is a member of​​​‌ the editorial board of​ Interstice.

9.3.2 Participation in​‌ Live events

• Amine Boumaza​​ participated twice in the​​​‌ Le Procès du Robot​.
• Amine Boumaza gave​‌ a conference on collective​​ robotics to students from​​ prépa at Lycée Henri​​​‌ Poincaré of Nancy.
• Olivier‌ Buffet participated in the‌​‌ robotic animation for elementary​​ students, during "fête de​​​‌ la science 2025", Univ.‌ Lorraine.
• Vincent Thomas participated‌​‌ in the robotic animation​​ for elementary students, during​​​‌ "fête de la science‌ 2025", Univ. Lorraine.

International journals

• 3‌ articleK.Kévin Bouillet‌​‌, S.Sophie Lemonnier​​, F.Fabien Clanché​​​‌ and G. C.Gérome‌ C Gauchard. Effects‌​‌ of pace on productivity​​ and physical and mental​​​‌ workloads in a Human-Cobot‌ Collaboration.International Journal‌​‌ of Occupational Safety and​​ Ergonomics311January​​​‌ 2025, 328-336HAL‌DOI
• 4 articleJ.‌​‌Jessica Colombel, V.​​Vincent Thomas, F.​​​‌Francis Colas and P.‌Pauline Maurice. Identifying‌​‌ human movement strategies for​​ human-robot collaboration.Multidisciplinary​​​‌ Biomechanics Journal49th congress‌ of the Société de‌​‌ BiomécaniqueJanuary 2025HAL​​DOIback to text​​​‌
• 5 articleS.Salomé‌ Lepers, S.Sophie‌​‌ Lemonnier, V.Vincent​​ Thomas and O.Olivier​​​‌ Buffet. How to‌ Exhibit More Predictable Behaviors‌​‌.The European Journal​​ on Artificial IntelligenceSeptember​​​‌ 2025HALDOIback‌ to text

International peer-reviewed‌​‌ conferences

• 6 inproceedingsR.​​Raphaël Boige, A.​​​‌Amine Boumaza and B.‌Bruno Scherrer. AlphaBeta‌​‌ is not as good​​ as you think: a​​​‌ simple class of synthetic‌ games for a better‌​‌ analysis of deterministic game-solving​​ algorithms.Advances in​​​‌ Neural Information Processing Systems‌ 39 (NeurIPS 2025)The‌​‌ Thirty-ninth Annual Conference on​​ Neural Information Processing Systems​​​‌San Diego, United States‌December 2025HALback‌​‌ to text
• 7 inproceedings​​K.Kevin Bouillet,​​​‌ S.Sophie Lemonnier,‌ E.Elisa Mathieu,‌​‌ L.Lien Wioland,​​ A.Alaa Hassan and​​​‌ M.Michaël Benedic.‌ Strategies for introducing cobots‌​‌ into SMEs: a self-assessment​​ tool for company maturity​​​‌.IHM'25 - 36e‌ Conférence Internationale Francophone sur‌​‌ l'Interaction Humain-MachineIHM’25 :​​ Actes étendus de la​​​‌ 36ème conférence Francophone sur‌ l’Interaction Humain-MachineToulouse, France‌​‌November 2025HAL
• 8​​ inproceedingsE.Erwan Escudie​​​‌, M.Matthia Sabatelli‌, O.Olivier Buffet‌​‌ and J.Jilles Dibangoye​​. ε-Optimally Solving Two-Player​​​‌ Zero-Sum POSGs.Advances‌ in Neural Information Processing‌​‌ SystemsAnnual Conference on​​ Neural Information Processing Systems​​​‌39San Diego (Californie‌ - EU) et Mexico‌​‌ City (Mexique), United States​​December 2025HALback​​​‌ to text
• 9 inproceedings‌G.Geoffrey Laforest,‌​‌ O.Olivier Buffet,​​ A.Alexandre Niveau and​​​‌ B.Bruno Zanuttini.‌ Post-Hoc Interpretation of POMDP‌​‌ Policies.Proc. 28th​​​‌ European Conference on Artificial​ Intelligence (ECAI 2025)28th​‌ European Conference on Artificial​​ Intelligence (ECAI 2025)Bologna​​​‌ (ITALY), ItalyIOS Press​October 2025HALback​‌ to text
• 10 inproceedings​​S.Salomé Lepers,​​​‌ V.Vincent Thomas and​ O.Olivier Buffet.​‌ Observer-Aware Probabilistic Planning under​​ Partial Observability.Proceedings​​​‌ of the 24th International​ Conference on Autonomous Agents​‌ and Multiagent Systems (AAMAS​​ 2025)AAMAS 2025 -​​​‌ 24th International Conference on​ Autonomous Agents and Multiagent​‌ SystemsDetroit (Michigan), United​​ StatesMay 2025HAL​​​‌back to text
• 11​ inproceedingsY.Yang You​‌, V.Vincent Thomas​​, F.Francis Colas​​​‌, R.Robert Skilton​ and O.Olivier Buffet​‌. Online Robust Robot​​ Planning for Human-Robot Collaboration​​​‌.Lecture Notes in​ Artificial IntelligenceTAROS 2024​‌ - 25th Toward Autonomous​​ Robotic Systems Conference15052​​​‌Lecture Notes in Computer​ ScienceLondres, United Kingdom​‌Springer Nature SwitzerlandDecember​​ 2025, 11HAL​​​‌DOIback to text​
• 12 inproceedingsY.Yang​‌ You, V.Vincent​​ Thomas, A.Alex​​​‌ Schutz, R.Robert​ Skilton, N.Nick​‌ Hawes and O.Olivier​​ Buffet. Partially Observable​​​‌ Monte-Carlo Graph Search.​Proceedings of the International​‌ Conference on Automated Planning​​ and SchedulingInternational Conference​​​‌ on Automated Planning and​ Scheduling351Melbourne,​‌ AustraliaSeptember 2025,​​ 279-287HALDOIback​​​‌ to text

Conferences without​ proceedings

• 13 inproceedingsK.​‌Kévin Bouillet, M.​​Muriel Lombard, A.​​​‌Alex Gabriel, A.​Alaa Hassan and S.​‌Sophie Lemonnier. Using​​ an augmented reality device​​​‌ to learn a collaborative​ task with a cobot​‌.La XR en​​ mouvementLimoges, FranceJune​​​‌ 2025HAL
• 14 inproceedings​G.Geoffrey Laforest,​‌ O.Olivier Buffet,​​ A.Alexandre Niveau and​​​‌ B.Bruno Zanuttini.​ Post-Hoc Interpretation of POMDP​‌ Policies.Actes des​​ Journées d'Intelligence Artificielle Fondamentale​​​‌ (JIAF 2025) et des​ Journées Francophones sur la​‌ Planification, la Décision et​​ l’Apprentissage pour la conduite​​​‌ de systèmes (JFPDA 2025)​Dijon, FranceMay 2025​‌HALback to text​​

Reports & preprints

• 15​​​‌ reportG.Geoffrey Laforest​, O.Olivier Buffet​‌, A.Alexandre Niveau​​ and B.Bruno Zanuttini​​​‌. Post-Hoc Interpretation of​ POMDP Policies (extended version)​‌.Université de Caen​​ NormandieAugust 2025HAL​​​‌back to text
• 16​ miscS.Salomé Lepers​‌, V.Vincent Thomas​​ and O.Olivier Buffet​​​‌. Observer-Aware Probabilistic Planning​ Under Partial Observability.​‌November 2025HALDOI​​back to text

Other​​​‌ scientific publications

• 17 inproceedings​K.Kévin Bouillet and​‌ S.Sophie Lemonnier.​​ Une méthode quantificative d’évaluation​​​‌ de la qualité des​ interactions entre un humain​‌ et un cobot..​​ModACT 2025 -3​​​‌1Paris, FranceMay​ 2025HAL